Nearshore Engineering Economics 2026: Rates, AI & Delivery

Key takeaways

→The typical software development vendor worldwide bills $30–$49 an hour. Only about 10% bill above $100.³
→Fully loaded nearshore engineering cost in Latin America runs $65,000–$72,000 a year against $165,000–$175,000 for a comparable US hire. In Latin America, the junior-to-staff gap is about $65,000 while the cheapest-to-most-expensive-country gap is about $10,000. Within the region, seniority mix matters far more than country selection.
→AI hasn't broadly pushed vendor pricing down. About 90% of vendors held or raised rates in early 2026, and GoodFirms' published project bands imply AI-powered builds can carry a roughly 1.25 to 1.7 times premium in comparable categories.
→Code output rose far more than shipped software. Across 100,000+ developers, AI coding agents pushed lines of code up several hundred percent while production releases moved 10 to 20%.
→A new offshore or nearshore team using AI heavily on an unfamiliar codebase might look fast early, but may also be building a weaker mental model of your system than a slower team would.
→Nearshore, while more expensive than offshore delivery, offers a valuable time-zone overlap. In an AI-heavy pipeline, that overlap improves review, integration, and release latency—so model the collaboration cycle, not just the hourly rate.
→The AI specialist premium is 10 to 20% at the same seniority level, not the 40 to 56% that vendors commonly quote. The larger numbers compare different seniority mixes.
→If you do one thing before signing anything: make your partner measure an idea-to-production baseline on your own work before the engagement starts. Without it, no productivity claim either of you makes later can be checked.

What software development costs per hour in 2026

To get a real sense of pricing, we looked at two independent surveys. These are more reliable than vendor tier tables because they show what companies are actually charging, not just what they want to charge.

Rate distribution across the vendor population

Hourly rate band	Share of firms (Techreviewer)	Share of firms (GoodFirms)
Under $20	8.7%	12.5%
$20–$29	22.1%	56.3% (reported as $20–$50)
$30–$49	37.8% — largest single band	—
$50–$99	22.0%	12.5% (reported as $50–$100)
$100–$149	6.3%	10.9% (reported as $100–$250)
$150–$199	3.2%	—
Above $250	—	7.8%

Techreviewer: 127 software development providers, fielded January–February 2026, published March 28, 2026. GoodFirms: 100+ global software companies, fielded September–October 2025, published March 17, 2026.

The overwhelming majority of software development vendors bill under $100 an hour—90.5% according to Techreviewer's numbers, 81.3% according to GoodFirms'. Beyond that, the two surveys stop lining up: GoodFirms finds 7.8% of firms above $250, while Techreviewer's top bracket ends at $150–$199 with 3.2%. Neither found a distinct population above $300.

Rates by firm class

The following tiers come from vendor-published positioning tables, not independent market surveys. It's best to read them as positioning rather than measurement. Since the numbers don't always line up, we've shown both major sources rather than averaging them into a cleaner number that the evidence does not support.

Firm class	RaftLabs (May 2026)	Simply.Coach (Feb 2026, upd. May 2026)
Enterprise consultancy or Big Four	$300–$500	$250–$850+
Large firm, 250–1,000 people	not broken out	$200–$300
Mid-market custom development	$175–$300	$100–$200 (50–250 people)
Boutique or small agency	$100–$250	$75–$175 (1–50 people)
Freelance, established	$75–$150	$80–$150

RaftLabs, May 29, 2026⁴; Simply.Coach, February 2026, updated May 2026.⁵Both vendor-published, neither with a disclosed methodology.

A note on freelancer pricing. Platform-sourced freelancers bill well below the "established freelance" tier: $15–$35 for entry-level front-end, full-stack, or JavaScript work, $20–$38 for intermediate React, $40–$100 for DevOps, and $75–$150 or more for specialized AI and consulting work (Upwork first-party rates, May 12, 2026). Nearshore and offshore firms sit below that again, in the regional section below.

A note on enterprise pricing. While earlier versions of FullStack's price guide claimed that enterprise rates reach $900 an hour, we've since removed that figure because we can no longer defend it. As far as we're aware, no major consultancy publishes a commercial rate card, with the only primary consultancy rate card we could obtain being Accenture's published US federal schedule. This uses a fiscal 2012 base year escalated at roughly 3% per year, and federal rates typically run below commercial ones.

Engagements above $500 an hour can happen. These are typically negotiated in rooms, and not listed on websites. As such, FullStack can't confidently publish numbers on these larger projects.

Pricing by Seniority

A common mistake is comparing a solo contractor's rate to an agency's rate. An agency usually costs 1.5 to 2.5 times more, with that extra cost covering necessities like project management, QA, and insurance that your project won't stop if one engineer quits. A solo freelancer is cheaper up until they aren't available when you need them.

Group 1: Solo contractor rates

Region	Junior	Senior	Year-over-year change
Latin America	$33–$45	$60–$75	−7.1%
Europe	$31–$39	$64–$76	−4.4%
Asia	$24–$31	$31–$41	approximately −8%

Accelerance blog summary of its 2026 Global Software Outsourcing Rates guide, published November 24, 2025, from a survey of 100+ firms.⁶The underlying guide is gated; these figures come from the published summary.

A second dataset built from signed contracts rather than survey answers lands in similar territory and adds seniority:

Seniority	Experience	Rate vs. senior	Share of contracts
Mid-level	3–5 years	28–45% below	not reported
Senior	5–8 years	baseline	49%
Lead / strong senior	8+ years	46–89% premium	13%

Lemon.io Salary Report 2026, published April 25, 2026: 2,500+ signed contracts, January 2024 to April 2026, 71+ countries. To their credit, they disclose that contractor rates run 30 to 50% above equivalent salaries, that their platform skews senior, and that Brazil alone is 20% of their contracts.⁷

Group 2: Agency rates by region

Region	Hourly rate band
United States	$150–$300
United Kingdom and Ireland	$130–$250
Canada and Australia	$100–$200
Western Europe	$100–$200
Eastern Europe	$50–$120
Latin America	$40–$100
India	$25–$75
Southeast Asia	$25–$65

RaftLabs, May 29, 2026. Vendor-published, no disclosed methodology — read as positioning, not measurement. Clutch's directory data, updated August 16–17, 2026, gives the US at $50–$99 on one pricing page and $100–$149 on another updated within a day of it. That inconsistency is why this guide doesn't quote a single "Clutch US rate."⁸

One caveat about direction of travel.Accelerance says rates are falling 4 to 8%. Techreviewer, fielding three months later, says roughly 90% of firms held rates flat or raised them. Both can be right: Accelerance is measuring outsourced delivery rates, where a shift toward cheaper geographies and more junior engineers pulls the average down without anybody cutting a price. Techreviewer asked firms what they did to their own rate card. If you're sitting across a table from a vendor, the useful read is this: list prices aren't falling, but the blend you get offered might be.

Where nearshore differs from offshore

Most comparisons ignore the practical differences between these two, but they change how you'll work day-to-day.

In Latin America, seniority matters more than the country. Salaries across the region are fairly close, with only about a $10,000 difference from the most expensive country to the cheapest. However, the gap between a junior and a staff engineer is $65,000. If you're hiring in LatAm, focus on getting the right mix of experience rather than the specific country.

Nearshore isn’t always the cheapest option. Senior talent in Latin America costs $60–$75 an hour, while Asian talent is around $31–$41. If you're looking for the most affordable option, then offshore teams in Asia or India will usually be the best pick.

What you’re really paying for with nearshore is time zone overlap. AI tools generate code fast, but reviewing and testing that code still requires human input. A team that works the same hours as your reviewers will move much faster than a team that only has one hour of overlap. It's up to you to decide whether that faster feedback loop is worth the higher rate.

The more useful comparison to model is: rate × seniority mix × required overlap × review latency, not rate alone.

Why rates don't tell the whole story anymore

An hourly rate only tells you the price of effort. What you actually want is working code in production. Over the last few years, AI has made it much easier to write code, but that doesn't mean it's easier to finish projects.

More code doesn't mean more software

The best-identified study on this tracked more than 100,000 GitHub developers from May 2022 to May 2026, using a matched design that compares every AI adopter against a near-identical developer a year earlier, cross-referenced against confidential Microsoft GitHub Copilot subscription telemetry from April to December 2025.⁹

Outcome measured	Autocomplete tools	Synchronous agents	Asynchronous agents
Lines of code	+228.2%	+741.3%	+658.3%
Commits	+35.9%	+109.1%	+33.6%
Pull requests	+11.0%	+65.5%	+71.8%
Releases	+10.2%	+20.3%	not applicable

Demirer, Musolff, and Yang, “Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools”, NBER Working Paper 35275, 2026.¹⁰

Researchers found that as you move from writing code to shipping it, the benefits of AI start to disappear. You can't just generate your way to a finished product; you still need humans to verify everything. Additionally, while more software is being built, people aren't necessarily using it more—total engagement with new apps is actually flat or down across major app stores.

The "one-third" rule

Writing and testing code account for about 25 to 35% of the time from initial idea to product launch, according to Bain's 2025 analysis. The rest is specification, architecture, review, integration testing, security verification, and deployment—work that AI has not fully automated.¹¹

If AI can only accelerate a third of the process, then even if code generation became instant and free, your total speedup caps out around 50 percent. It explains a common failure in AI business cases: the model assumes the acceleration applies to the whole lifecycle when it applies to one slice of it.

Bain's own read, from teams that rolled out AI assistants versus teams that redesigned the process end to end, supports this: assistant rollouts alone produced 10 to 15% gains, full process change reached 25 to 30%. As Bain notes, "the time saved often isn't redirected toward higher-value work, so even the modest gains that have been made have not translated into positive returns".

Test any vendor's savings claim against the one-third ceiling. If they're promising more than it allows, either they're changing your process too, which is a much bigger engagement than they're quoting, or they haven't done the arithmetic.

Review becomes the bottleneck, and then it breaks

When code generation gets cheap, and review doesn't, that pressure moves downstream. Telemetry across 22,000 developers and 4,000+ teams, comparing each organization against itself at its lowest and highest AI adoption, shows exactly where it lands. (Faros)

Pipeline measure	Change	What it means for you
Average pull request size	+51.3%	Bigger changes, harder to review properly
Median time in review	+441.5%	Review has become the binding constraint
Changes merged with no review	+31.3%	The corner that gets cut when queues overflow
Code churn	+861.0%	Code written, then thrown away and rewritten
Production incidents per change	+242.7%	Instability reaching your customers
Epics completed	+66.2%	The roadmap looks like it's accelerating
Weekly production deployments	−11.7%	Actual delivery went backwards

Faros AI, “AI Engineering Report 2026: The Acceleration Whiplash”. These are within-organization before-and-after correlations, not causal estimates, and Faros sells engineering-intelligence software.¹²

If you look at these two rows, you'll see that projects might look like they are moving faster on paper (more "epics" completed), while actual deployments are slower. If you're just watching the roadmap, you might think you're winning while less software is actually reaching your users.

Be mindful of the review bottleneck. When AI increases the rate at which code arrives, review, integration, security, and release can become the constraint. If review capacity does not scale with output, the apparent savings show up upstream first while reliability and maintenance costs can surface later, downstream. Before you sign, name who reviews the work, the expected volume, the turnaround commitment, and what happens when the queue backs up.

Two more findings support this. Across 623 million code changes from 2023 to 2026, copy-pasted code rose 41 to 15.7% of changed lines while structured refactoring collapsed from 21% of changes to 3.8 percent.¹³

Additionally, across 150+ models and 80 standardized tasks, 45% of generation tasks produced code containing a known security flaw, with syntax correctness above 95% and security pass rates near 55%.¹⁴Code that compiles cleanly and fails a security review is the most expensive kind, because it passes the checks a non-technical reviewer knows how to run.

Verification costs real time

In Stack Overflow's 2025 Developer Survey, 66% of respondents cited “AI solutions that are almost right, but not quite” as the top source of frustration, while 45% said debugging AI-generated code is more time-consuming. 84% of respondents said they use or plan to use AI tools in their development process. On accuracy, 46% distrust AI output, 33% trust it, and only 3% report highly trusting it.¹⁵

Objective, unbiased reporting is elusive

In a controlled trial of 16 experienced open-source maintainers completing 246 tasks on codebases over a million lines, developers believed AI made them about 20% faster. The clock said they were 19% slower—a 39-point gap between perception and measurement.¹⁶

That study has since been walked back by its own authors, and the walk-back matters as much as the original. Expanded in early 2026 to 57 developers and 800+ tasks, the original cohort came in at −18% with a confidence interval spanning −38 to +9%, while newly recruited developers came in at −4%.¹⁷

The researchers documented serious selection bias: the most productive AI-enabled engineers declined to take part because the design required working half the time without AI, participant compensation had been cut from $150 to $50 an hour, and multi-agent workflows couldn't be captured by single-task time logs at all. While you shouldn't cite “19%” slower as fact, it's worth keeping the 39-point perception gap. It's the reason a vendor's confident report of AI-driven speedup, offered without measurement, is worth nothing.

The evidence shows that the jury is still out

The maintainability and defect evidence above is real. However, so is the evidence pointing the other way. Three field experiments covering 4,867 developers found pull request approval rates up about 10% and no evidence that code quality decreases. In the 2025 DORA survey, 59% reported a positive impact on code quality.¹⁸The defensible conclusion isn't that AI degrades quality—it's that outcomes diverge sharply by organization, which puts the decisive variable inside your delivery system rather than inside the tool. Which means the real question is no longer “does this vendor use AI,” as they all do—it's whether their delivery system can absorb it, and whether yours can.

The AI premium, measured honestly

The premium for AI expertise is real and much smaller than the headline numbers suggest. Its size depends entirely on what you compare it against.

Comparison basis	Premium	Source
AI vs. non-AI peers at the same level	+6.2% entry, +11.9% engineer, +14.2% senior, +18.7% staff	Pin.com aggregation, May 2026 (vendor)
AI professionals vs. data scientists, cash compensation	+9–13%	Burtch Works 2025, n=866
AI/ML vs. base developer rate, Latin America	+12–15%	Mismo, March 2026 (vendor)
Annual salary growth, AI vs. general tech	+4.1% vs. +1.6%	Robert Half 2026, via Pin.com
Premium on AI skills in job postings	+56%	PwC 2025, via Pin.com

What holds up is 10 to 20% at the same level. The 40 to 56% figures in circulation aren't like-for-like. PwC's 56% measures a wage premium attached to AI skills appearing in job postings, which differs from two engineers of equal seniority. Compare an AI/ML engineer median against a full-stack developer median and you get about 37%, but most of that is AI/ML engineers simply being more senior on average.

FullStack Data Perspective

Comparing engineering roles against engineering roles in our own filled placements, the AI premium works out at about 13%. That sits inside the 10-to-20% range these third-party sources report. What’s interesting is that on widening the comparison to every role type, the premium reads about 15% at the median and 25% at the mean, because that mixes in architects and project managers whose rates are higher for reasons that have nothing to do with AI.

Another point to note is that what counts as AI work in our data is the skills a seat actually requires — large language models, AI agents, Model Context Protocol, agentic tooling, applied machine learning — not a job title with “AI” in it. If a vendor quotes you an AI premium, ask which of those two they’re pricing.

If you need a true AI specialist, the US-to-LatAm price difference is still notable:

Role	US base median	Latin America, all-in
Agentic AI engineer	$223,000	$134,000
RAG engineer	$192,000	$115,000
MLOps engineer	$190,000	$114,000
Generative AI engineer	$190,000	$114,000

Next Idea Tech, May 29, 2026, based on 467 US senior AI job postings from April–May 2026 plus placement pricing. US total compensation at well-funded companies runs 20 to 40% above base, reaching $350,000 or more.¹⁹

It's important not to confuse AI engineers with AI "trainers" or data annotators. Those roles are much cheaper ($15 to $75 an hour). If a proposal lists low rates for AI work, ask whether they're actually hiring engineers or just people to label data.

Engineering salaries by region

What people earn in the US

Measure	Figure	Period
Software developers, median hourly	$65.38	May 2025
Software developers, mean annual	$148,100	May 2025
Software developers, median annual	$133,080	May 2024
QA analysts and testers, median annual	$102,610	May 2024
Projected employment growth, 2024–34	+16% developers, +10% QA	—

Numbers courtesy of the US Bureau of Labor Statistics.²⁰ It's worth noting that $148,100 is the average salary, not the median.

United States medians by role

Role	US median annual salary
Engineering manager	$200,000
AI/ML engineer	$189,500
Cloud infrastructure engineer	$189,000
Architect	$180,000
Back-end developer	$175,000
Mobile developer	$170,000
DevOps engineer	$165,000
Data engineer	$150,000
Front-end developer	$145,000
Full-stack developer	$138,000

Numbers courtesy of theStack Overflow Developer Survey 2025, 49,000+ respondents.

Latin America

Seniority	Annual take-home
Junior	$40,000–$45,000
Mid-level	$50,000–$60,000
Senior	$65,000–$75,000
Principal	$80,000–$95,000
Staff	$90,000–$105,000

Numbers courtesy ofHowdy, 12,500+ compliant employment agreements across seven countries, published May 2026, updated August 13th, 2026.²¹

On the same source's figures, fully loaded Latin American engineering cost runs about $65,000–$72,000 against $165,000–$175,000 for a comparable US hire—a 60 to 65% difference in employer cost. A fully loaded figure and a take-home figure for the same engineer differ by well over a third, which is why people often confuse the two.

You might see higher numbers elsewhere—some datasets suggest a LatAm median over $100,000. That usually means they are looking at senior engineers working for well-funded US companies. Our rule is to always ask if a salary number is take-home pay or the total cost to the employer, as they aren't the same thing.

Eastern Europe

These are monthly gross rates in Euros.

Country	Junior	Mid	Senior
Poland (employment)	€1,700–€2,100	€2,800–€4,000	€4,000–€6,000
Poland (B2B contract)	€2,000–€2,500	€3,500–€5,000	€5,500–€8,000+
Czechia	€1,730–€2,200	€2,500–€3,500	€3,800–€4,930
Hungary	€1,800–€2,200	€2,400–€3,200	€3,300–€4,500
Romania (product or remote)	€1,200–€1,600	€2,400–€3,200	€4,000–€6,000+
Romania (local or outsourcing)	€1,000–€1,300	€1,600–€2,400	€2,800–€3,400
Bulgaria	€1,000–€1,400	€1,600–€2,300	€2,300–€3,500

Numbers courtesy of CEEhire, dated May 1st, 2026.²²

What affects AI use

AI doesn't work the same for everyone. Only about one in five digital transformation projects reach their goals, which is often due to the organizational culture around AI tools.²³By taking certain steps, you can not only improve your company's culture, but also help your teams get more value from the tools they use.

• A clear, communicated position on AI use: 15% of US employees strongly agree that their company has clearly explained its plan for using AI. Clarify, in writing, what is and isn't allowed when using AI tools.
• Healthy data ecosystems: AI is more useful when it can draw on accurate, connected internal data. Make sure the information your teams rely on is easy to find, up to date, and managed consistently.
• AI-accessible internal context: Can your AI tools securely access the code, documentation, and architecture information they need?
• Strong version control practices: Clear version history, regular commits, and reliable rollback options make it easier to review and reverse AI-generated changes.
• Working in small batches: Break work into smaller changes that are easier to review, test, and reverse. This helps teams catch problems before they become harder to untangle.
• A user-centred focus: AI should help teams solve real user problems, not simply produce more output. Keep user needs in view when deciding where and how to use it.
• Quality internal platforms: Give teams clear, automated paths for common work, with appropriate guardrails built in. That reduces the need for every team to create its own approach from scratch.

AI works best when teams have strong processes in place. Without the right framework, it may expose existing problems rather than help solve them. We recommend that if four or more of these areas need work, you should focus on your delivery system before hiring more people.

Two findings support this pattern. First, task selection matters enormously: a 2026 ROI analysis reports roughly 35–40% productivity gains on simple greenfield tasks but 10% or less on complex legacy code, and explicitly flags those figures as highly uncertain, so treat the spread as directional rather than precise.²⁴

Second, the distribution of gains by seniority is uneven: across three field experiments covering 4,867 developers, junior and short‑tenure engineers saw gains of about 21–40%, while senior and long‑tenure engineers saw only 7–16%.²⁵

A warning for new teams. In Anthropic's January 2026 randomized controlled trial, 52 mostly junior engineers learned an unfamiliar Python library with or without AI assistance. The AI-assisted group scored 17 percentage points lower on a subsequent mastery quiz, with the largest gap on debugging; participants who relied most heavily on AI scored below 40% on average. While the study is small and measures immediate comprehension, not long-term skill development, it's a useful onboarding warning: early output can improve before system understanding does. FullStack's mitigation is to have the team explain the architecture, own incident postmortems, and show they can reason about the system without handing every debugging step to an agent.

Seven questions that tell you who AI is benefiting

The biggest question during negotiations is this: how do you know you're actually getting the efficiency you're paying for?

Almost every vendor uses AI now. The question is: are they pocketing the savings, or are they passing them on to you? Here's how to find out.

1. What was our baseline, and what is it now? Ask for idea-to-production cycle time on a defined set of work, measured at the start. If a vendor can't name a baseline, there's nothing to verify later, and self-reporting won't cover for it—in the best-known trial on the subject, developers reported a speedup the measured data didn't show.
2. What share of merged code on our project is AI-authored, and how do you know? If a partner can't measure code provenance or AI-tool use reliably, they can't tell you how AI is changing your delivery system. The point is not to police a percentage; it is to make the change observable.
3. What changed after you adopted AI? Ask whether the vendor's rates, team size, delivery timeline, or total project cost changed after it adopted AI tools. If nothing changed, the vendor might be keeping their efficiency gain. Though this isn't necessarily an issue, they should be transparent about it.
4. What has happened to your change failure rate and review time? Faster shipping with unchanged stability is the claim worth probing—industry telemetry shows review time up fivefold and incidents per change up 242.7% in high-adoption periods. (Faros)
5. Who reviews AI-generated code before it reaches our repository, and what is your unreviewed-merge rate? Unreviewed merges rose 31% in that same telemetry. This is the specific corner that gets cut when review becomes the bottleneck, and it gets cut quietly.
6. What do you charge for AI-specific work versus standard engineering, and why? AI builds carry a real premium over comparable general builds, which can be legitimate. However, a vendor charging an AI premium while claiming AI savings owes an explanation of which one applies where.

7. Will you put the measurement in the contract? A partner confident about their gains should be willing to agree to a measurable reporting mechanism; if not, ask why. This is the cheapest test on this list and one of the most revealing.

What else to put in the contract

After you've asked the questions above, put the answers that matter into the contract so they stay explicit after work begins. It's also worth discussing the six provisions we mentioned above. A good partner should be clear about what it can agree to, what it would change, and why.

1. A measured baseline before work starts. Measure idea‑to‑production lead time on defined, representative work units, using observed data rather than estimates. Without this baseline, no later claim about impact is verifiable by either party.
2. Code provenance and attribution. Maintain automated, commit‑level logs that distinguish machine‑authored code, human‑written logic, and subsequent deletions. This is often the main way for both sides to see whether generated code is accumulating unreviewed.
3. Explicit treatment of efficiency gains. If a partner adopts AI tooling and their blended rate, team size, and your timeline all stay the same, the productivity surplus is entirely theirs. That can be a defensible commercial position, but it should be a stated one. Negotiate a smaller squad or a compressed timeline rather than assuming a discount will arrive.
4. Downstream quality metrics in the service levels, not just output metrics. Review latency caps, change failure rate thresholds, and defect density limits. A contract that rewards volume without constraining quality pays a vendor to transfer maintenance cost to you.
5. An unreviewed-merge ceiling. Unreviewed merges—human or agentic—rose 31% in high-adoption telemetry. Counter this by setting a hard threshold and report against it monthly. (Faros)
6. AI premiums separated from AI theatre. Dedicated AI engineering roles carry a 10 to 20% like-for-like premium, and specialist work—retrieval systems, evaluation pipelines, model routing, guardrails—is genuine scope.²⁶Standard development using an off-the-shelf coding assistant is not. Make a partner state which is which, and price accordingly.

The first 90 days

Window	What must happen	Why
Before day 1	Idea-to-production cycle time measured on defined, representative work. This should be written down and signed off by both sides.	This is the only thing that makes any later claim checkable. Skip this, and you've lost the ability to verify.
	Named team roster with seniority and years of experience per person.	Seniority mix is the largest budget variable and the least disclosed.
	Review capacity named and resourced on your side.	See the review-capacity trap above.
	The seven-capability readiness check completed honestly.	Fix what you can, and know what you're carrying.
Days 1–30	Team ships something small to production, end-to-end.	Tests the whole pipeline, not the coding step. Nearly every engagement discovers its real constraint here.
	AI-authorship attribution turned on at the commit level.	Creates a measurable record of how AI use is changing the codebase instead of relying on self-report. You can't manage what you can't see.
	Architecture walkthrough presented by the team.	Verifies mental models, not just output.
Days 31–60	First measured comparison against the day-zero baseline.	Direction of travel, not a verdict.
	Review latency and unreviewed-merge rate reported.	The two leading indicators of the review-capacity trap.
	First incident postmortem written by the team.	If they can't diagnose it, tooling is masking a comprehension gap.
Days 61–90	Full metric set: cycle time, review latency, change failure rate, deployment frequency, unreviewed-merge rate.	Enough data to make a real renewal decision.
	Decision point: expand, hold, or stop.	Made against measurement, not against how the relationship feels.

This checklist is less about the team you're hiring, and more about your own ability to see what they are doing. That's often where most projects are won or lost.

What projects cost, and what happens after launch

What typical projects cost

Project class	Cost range	Share of projects
Small or MVP	Under $30,000	14.1%
Medium	$30,000–$100,000	65.7%
Large-scale	$100,000–$200,000	14.1%
Enterprise	$200,000+	6.3%

Numbers courtesy ofGoodFirms, 100+ global software companies, fielded September–October 2025.

The average project costs about $132,000 and runs for 13 months, but that average is pulled upward by a small number of much larger engagements. Most projects fall between $10,000 and $49,000. When budgeting, it's best to reference that common range over the average.

Why AI builds are usually more expensive.

Build type	Small or MVP	Medium	Large or enterprise
General software build	Under $30,000	$30,000–$100,000	$100,000–$200,000+
AI-powered build	$50,000–$125,000	$125,000–$250,000	$250,000+

GoodFirms' published 2026 project bands imply a roughly 1.25 to 1.7 times premium for AI-powered builds in comparable categories. That premium can reflect higher-priced specialist talent, evaluation and guardrail work, and recurring inference costs. However, it doesn't mean every AI project costs 1.25 to 1.7 times more; scope and architecture matter. It's also worth noting that most surveyed firms held or raised their rate cards in early 2026.

Maintenance: what we had wrong

While we used to claim that maintenance costs 15 to 25% of the original build, we've since realized we can't trace that range to a reliable source. From what we've gathered, it seems to be an industry rule of thumb that gets repeated without much evidence.

We also couldn't find enough reliable data to replace it with a new, universal percentage. The cost of maintenance will ultimately vary based on the type of app, how much work it needs, and whether the work includes improvements or new features. For some apps, that might mean a few bug fixes and keeping systems up to date, while others might require a more thorough—and more expensive—overhaul to improve performance, strengthen security, or add new features.

Additionally, moving data or connecting an app to other systems can also drive up the cost. These projects should be estimated separately because the work will depend on the systems involved, the condition of the data, and the testing required.

Things that are still unclear.

There are still a few questions we don't have answers to. Here's what we don't know as of August 2026.

We still don't know the real "cost per feature." We know what teams pay in salaries and hourly rates, but there isn't a credible published study showing what it actually costs to ship a finished piece of software using AI. Any specific “cost per feature” number is closer to a rough guess than a measured result. However, we're currently working on figuring that out for our own teams.

The headline ROI models are explicitly illustrative. The most-cited 2026 return figure for AI-assisted development, 39% first-year ROI with an eight-month payback, comes from DORA's RoI of AI calculator. This is a modelled scenario its own authors describe as "a highly uncertain estimate, intended to spark a conversation, not as a rigid mathematical formula." It omits net present value and discounting. It's being passed around boardrooms as a forecast.

Why “95% of AI pilots fail” is misleading. According to a study by MIT NANDA, 95% of organizations in its sample saw no measurable financial return from custom enterprise generative AI, based on 52 interviews and 153 survey responses.²⁷

The underlying study found 95% of organizations getting zero measurable financial return on custom enterprise generative AI. Among organizations that actually ran a structured pilot, roughly 25% reached production. It rests on 52 interviews and 153 survey responses, is not peer-reviewed, self-declares selection bias, and studies enterprise generative AI broadly rather than software engineering specifically.

AI use is widespread; enterprise-level financial impact is not.McKinsey's 2025 Global Survey found that 88% of organizations regularly use AI in at least one business function, while only 39% report any EBIT impact at the enterprise level. BCG AI Radar 2026 found that 94% of organizations planned to continue investing in AI even if it didn't deliver returns in 2026.²⁸Conviction is ahead of measurement. That does not mean the investment is wrong; it means the ROI case should be measured rather than assumed.

Quality is not settled against AI. AI's effect on quality still isn't clear-cut. The maintainability and defect problems we cited earlier are real, but so are the studies showing AI helping quality. What actually separates the two outcomes isn't the tool—it's the organization using it, and specifically the delivery practices they have in place around it.

What to measure instead of what to pay

Rates are easy to compare but they don't tell you much. Our stance is simple: if AI makes delivery faster, you should be able to see that value—either in a smaller team, a faster timeline, or a lower total price. If none of those things change, the vendor is just making a higher profit. That's their right, but it shouldn't be a secret.

Four things separate vendors in 2026, and you can check all four before you sign anything.

Will they measure a baseline? Without a starting point, you can't prove you're saving money later. Ask them exactly what they measure and how often you'll see the reports.

Can they prove their hiring process works? It is easier than it used to be to produce a polished résumé or take-home exercise with AI. Ask how the partner actually vets people. We share reasoning and communication signals with our clients so they can inspect the evidence for themselves. If a vendor will not let you inspect the evidence, treat that as a diligence gap.

Is their process a "secret"? Calling a method "proprietary" is not evidence. Ask for enough detail that your engineers can evaluate the workflow, controls, and measurement. A method should be specific enough to challenge, even if some implementation details remain confidential.

What do you own when the project is done? You should own your code, data and configuration, evaluations, operating documentation, and any custom model assets contractually created for you. If third-party models are involved, make the access, portability, and substitution rights explicit. And the engagement should leave your team more capable than when it started — make that an explicit deliverable, not an assumption.

If your problem is that your delivery lifecycle is slower than your tooling should allow, that's what AI PDLC addresses — tracing real features from idea to production, scoring the organization against a maturity model, and locking a baseline before anything changes. If your problem is that high-volume routine AI work is running on frontier-model pricing, Specialized Models is where the cost argument actually holds. If you can't attribute your AI spend to a team or a feature, AI Gateway & Model Routing is where that attribution gets built.

If you fail most of the readiness checks we listed earlier, our advice is to fix your internal delivery process before you hire a new team. It's better to solve the bottleneck first than to pay for extra capacity that just gets stuck in a queue.

Where to start

Before committing to a larger transformation, ask your provider what a typical project costs and how long it takes to deliver. Those numbers give you a clear starting point and something concrete to measure against.

If you're interested in comparing raw market pricing first, use FullStack's 2026 Software Development Price Guide. If your engineering lifecycle is slower than your tooling suggests it should be, the AI PDLC engagement starts by measuring how work actually moves through your system today, not by leading with projections about future gains.

The decision checklist

Take this into the vendor conversation.

On cost

Is the quoted figure a contractor rate or an agency rate? (They differ by 1.5 to 2.5 times)

Is the salary basis take-home, gross, or fully loaded? (They differ by more than a third)

What is the named seniority of every person on the proposed team?

What happens to the blend if someone leaves, and who carries that cost?

On AI

What share of merged code on our project will be machine-authored, and how will you know?

What are you charging as an AI premium, and for which specific work?

Has your blended rate, squad size, or delivery timeline changed since you adopted AI tooling? If none of the three moved, where did the efficiency go?

What has happened to your change failure rate and review time?

On measurement

Will you measure an idea-to-production baseline on our work before we start?

Will you put that measurement in the contract and report against it?

What is your unreviewed-merge rate, and will you commit to a ceiling?

On our side

Have we resourced review capacity to match the code volume we're about to add?

How many of the seven readiness capabilities do we actually pass?

Who on our side owns the 90-day decision, and against what numbers?

On the end of the engagement

What do we own when it ends — code, data and configuration, evaluations, operating documentation, any custom model assets created for us, and a team more capable than when we started? If third-party models are involved, are access, portability, and substitution rights explicit?

Common FAQs

What does software development cost per hour in 2026?+

Between 80 and 90 percent of software development vendors worldwide bill under $100 an hour, depending on which 2026 survey you use, and the largest single Techreviewer band is $30 to $49. Vendor-published firm-class tables put enterprise consultancies at roughly $250 to $850+ an hour, large firms at about $200 to $300, mid-market custom development at roughly $100 to $300, boutique agencies at about $75 to $250, and established freelancers at $75 to $150. The geography table puts agency rates at $40 to $100 in Latin America, $25 to $75 in India, and $25 to $65 in Southeast Asia. Agency billing rates can run roughly 1.5 to 2.5 times individual-contractor rates in the same country because they include delivery management, QA, architecture, employment costs, and bench capacity.

Has AI made software development cheaper?+

Not broadly in vendor pricing, as of August 2026. GoodFirms' published project bands put small AI-powered builds at $50,000 to $125,000 and medium builds at $125,000 to $250,000, above the published bands for general software projects. About 90 percent of surveyed firms held or raised their rate cards in early 2026. Coding output has risen sharply, but releases have risen far less — across more than 100,000 developers, synchronous AI agents raised lines of code 741 percent and releases 20.3 percent. AI can still reduce delivery cost in a specific organization, but you need a baseline to prove that the gain reached the buyer rather than stopping upstream.

How much does it cost to hire a software developer in Latin America?+

Take-home pay runs about $40,000 to $45,000 for junior engineers, $50,000 to $60,000 mid-level, $65,000 to $75,000 senior, $80,000 to $95,000 principal, and $90,000 to $105,000 staff, with a regional average around $57,000. Fully loaded employer cost across the region runs roughly $65,000 to $72,000, against $165,000 to $175,000 for a comparable US hire. Always confirm whether a quoted figure is take-home, gross, or fully loaded — published estimates for the same region differ by more than 80 percent depending on the basis.

Does nearshore staffing give better timezone overlap than offshore, and why does it matter?+

Yes, materially — but not because nearshore is always cheaper. The 2026 contractor data in this guide puts senior Latin American talent at $60–$75 an hour versus $31–$41 in Asia, so offshore can win on raw rate. Inside Latin America, however, country-to-country compensation varies by only about $10,000 while junior-to-staff compensation varies by about $65,000. Nearshore's distinct advantage is timezone overlap, which matters because review latency can become the binding constraint in an AI-heavy delivery pipeline. A team with six hours of daily overlap with your reviewers has a materially different feedback loop from a team with one hour, even at identical engineering quality. Price the overlap and seniority mix alongside the rate.

What should be in an offshore or nearshore engineering contract?+

Six provisions are worth negotiating: a measured idea-to-production baseline before work starts; code-provenance or AI-use measurement appropriate to the repository; an explicit, written position on who keeps the efficiency gain if the vendor's rate, team size, and your timeline do not change; downstream quality metrics such as review latency, change failure rate, and defect density alongside output metrics; a defined policy and reporting threshold for unreviewed merges; and AI premiums separated from standard development using off-the-shelf assistants. A strong partner should be able to tell you which it will accept, where it proposes an alternative, and why.

How much do AI engineers cost compared with regular software engineers?+

About 10 to 20 percent more at the same seniority level. Larger figures in circulation are not like-for-like: a frequently cited 56 percent premium measures the wage uplift attached to AI skills in job postings, not two engineers of equal seniority. In absolute terms, US AI/ML engineer median salary is around $189,500 against $175,000 for a back-end developer and $138,000 for a full-stack developer. Senior specialist roles run higher — agentic AI engineers at a US base median of $223,000, against roughly $134,000 all-in for the same role in Latin America.

How do I check whether a development partner is passing AI efficiency on to me?+

Ask seven questions: what baseline was measured at the start; what share of merged code is AI-authored and how they know; whether their blended rate, team size, or your timeline has changed since they adopted AI tooling; what has happened to their change failure rate and review time; who reviews AI-generated code and what their unreviewed-merge rate is; what they charge for AI-specific work versus standard engineering and why; and whether they'll put the measurement in the contract. Then put the answers you like into the contract itself — a vendor confident in the gain will agree to report against a baseline.

How should I budget for a software project?+

Define and prioritize scope first, then choose the engagement model. Estimate against the modal band for your project class rather than the average, which is skewed high by large outliers — a typical engagement falls between $10,000 and $49,000 in verified client reviews, against an average of $132,480. Add 10 to 25 percent for scope movement, and 20 to 30 percent if significant integration or data migration is involved. Then budget 10 to 20 percent of build cost annually for the lifecycle after launch. Finally, agree what gets measured before work starts. Without a baseline you can't tell later whether you got what you paid for.

Methodology and sources

Every figure here is current as of August 2026 and attributed inline to its publishing organization and date. In the published web version, each named source should link to the original study, dataset, survey, or first-party publication wherever one is publicly available; vendor summaries should remain labeled as vendor-sourced rather than being presented as independent measurement. Where two credible sources disagree, both are shown, and the difference is explained rather than resolved in favor of the more flattering number. Where a widely circulated figure has no traceable primary source — the 15-to-25% maintenance rule being the clearest example—the guide says so and uses the sourced alternative instead. Vendor-published figures are marked as vendor-published. Correlational findings are not described as causal. This report also includes perspectives from our own data, and we have clearly pointed out where we agree or diverge from publicly quoted numbers.

Rate and salary data draws on: Techreviewer (127 providers, fielded January–February 2026); GoodFirms (100+ software companies, fielded September–October 2025); the published summary of the Accelerance 2026 Global Software Outsourcing Rates guide (November 2025); Lemon.io Salary Report 2026 (2,500+ signed contracts); Howdy (12,500+ Latin American employment agreements, updated August 2026); Terminal 2026 Remote Software Engineer Salary Insights Report (July 2026); CEEhire (May 2026); Clutch directory and review data (August 2026); US Bureau of Labor Statistics Occupational Employment and Wage Statistics; Stack Overflow Developer Survey 2025; Gartner's 2026 IT spending forecast (July 2026); Deel State of Global Hiring 2026; Next Idea Tech (May 2026); and vendor-published tier tables from RaftLabs, Simply.Coach, Upwork, Pin.com, Burtch Works, Mismo, and FoundersBar, each marked as vendor-sourced where cited.

Delivery, productivity, and enterprise-AI evidence: Demirer, Musolff, and Yang, "Writing Code vs. Shipping Code" (2026); Bain Technology Report 2025; Faros AI Engineering Report 2026; GitClear "The Maintainability Gap"; Veracode Spring 2026 GenAI Code Security Update; DORA 2024–2025 reports and AI Capabilities Model; Cui, Demirer et al. in Management Science; METR 2025 study and February 2026 revision;Anthropic coding-skills study (Jan. 2026); McKinsey State of AI 2025; and BCG AI Radar 2026.

Next review: February 2027. If you find a figure here that has moved, or one we've got wrong, tell us and we'll fix it. Two of the corrections on this page came from someone doing exactly that.

Bibliography

GoodFirms. “Custom Software Development Cost Survey.” GoodFirms, https://www.goodfirms.co/resources/custom-software-development-cost-survey. Accessed 18 Aug. 2026.
Carle, Ben. Personal interview. Conducted by FullStack, 18 Aug. 2026.
Techreviewer. “Software Development Company Rates in 2026: Market Trends.” Techreviewer, 28 Mar. 2026, https://techreviewer.co/research/software-development-company-rates-in-2026. Accessed 18 Aug. 2026.
RaftLabs. “Custom Software Development Cost in 2026.” RaftLabs, 29 May 2026, https://www.raftlabs.com/blog/custom-software-development-cost. Accessed 19 Aug. 2026.
Simply.Coach. “IT Consulting Hourly Rates in the USA (2026).” Simply.Coach, 18 Feb. 2026, https://simply.coach/blog/it-consulting-rates-industry/. Accessed 19 Aug. 2026.
Accelerance. 2026 Global Software Development Rates & Trends Guide. Accelerance, 24 Nov. 2025, https://www.accelerance.com/2026-global-software-development-rates-trends-guide. Accessed 19 Aug. 2026.
Lemon.io. “Software Developer Salary Report 2026: Real Rates by Stack, Seniority & Region.” Lemon.io, 25 Apr. 2026, https://lemon.io/salary-report/. Accessed 19 Aug. 2026.
Upwork. “First-Party Software Development & Freelance Hourly Rates Data.” Upwork, 12 May 2026, https://www.upwork.com. Accessed 19 Aug. 2026.
Demirer, Mert, Leon Musolff, and Liyuan Yang. “Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools.” NBER Working Paper 35275, National Bureau of Economic Research, 2026, https://doi.org/10.3386/w35275.
Clutch. “Software Development Directory & Regional Pricing Data.” Clutch, 17 Aug. 2026, https://clutch.co. Accessed 19 Aug. 2026.
Demirer, Mert, et al. “Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools.” NBER Working Paper 35275, National Bureau of Economic Research, 2026, https://www.nber.org/papers/w35275.
Bain & Company. “From Pilots to Payoff: Generative AI in Software Development.” Technology Report 2025, Bain & Company, 23 Sept. 2025, https://www.bain.com/insights/from-pilots-to-payoff-generative-ai-in-software-development-technology-report-2025/.
Faros AI. “The Acceleration Whiplash: AI Engineering Report 2026.” Faros AI, 2026, https://www.faros.ai/research/ai-acceleration-whiplash. Accessed 19 Aug. 2026.
GitClear. “The Maintainability Gap: AI Code Quality in 2026.” GitClear, Jan. 2026, https://www.gitclear.com/the_ai_code_quality_maintainability_gap.
Veracode. “Spring 2026 GenAI Code Security Update: Despite Claims, AI Models Are Still Failing Security.” Veracode, 24 Mar. 2026, https://www.veracode.com/blog/spring-2026-genai-code-security-update.
Stack Overflow. “AI.” 2025 Stack Overflow Developer Survey, Stack Overflow, 2025, https://survey.stackoverflow.co/2025/ai. Accessed 19 Aug. 2026.
METR. “Measuring the Impact of Early‑2025 AI on Experienced Open‑Source Developers.” METR, 10 July 2025, https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/. Accessed 19 Aug. 2026.
METR. “We Are Changing Our Developer Productivity Experiment Design.” METR, 24 Feb. 2026, https://metr.org/blog/2026-02-24-uplift-update/. Accessed 19 Aug. 2026.
Next Idea Tech. “Nearshore AI Engineer Cost: 2026 US vs. LATAM Salary Data.” Next Idea Tech Blog, 29 May 2026, https://blog.nextideatech.com/nearshore-ai-engineer-cost-2026/. Accessed 20 Aug. 2026.
Salva, Ryan J. “How Are Developers Using AI? Inside Our 2025 DORA Report.” Google Blog, 23 Sept. 2025, https://blog.google/innovation-and-ai/technology/developers-tools/dora-report-2025/.
US Bureau of Labor Statistics. “Software Developers, Quality Assurance Analysts, and Testers.” Occupational Outlook Handbook, US Bureau of Labor Statistics, May 2025, https://www.bls.gov/ooh/computer-and-information-technology/software-developers.htm. Accessed 19 Aug. 2026.
CEEhire. “CEE IT Salaries in 2026: What Developers in Hungary, Poland, Romania, and the Region Actually Earn.” CEEhire, 1 May 2026, https://ceehire.com/blog/cee-it-salaries-in-2026-what-developers-in-hungary-poland-romania-and-the-region-actually-earn. Accessed 20 Aug. 2026.
Howdy. “2026 Latin America Software Developer Salaries: Verified Data from Howdy.” Howdy, 13 Aug. 2026, https://www.howdy.com/blog/latin-america-software-developer-salaries-2026. Accessed 19 Aug. 2026.
Ratanjee, Vibhas, and Ken Royal. “Your AI Strategy Will Fail Without a Culture That Supports It.” Gallup, 15 Aug. 2024, https://www.gallup.com/workplace/652727/strategy-fail-without-culture-supports.aspx. Accessed 19 Aug. 2026.
DORA. “ROI of AI-Assisted Software Development Calculator.” DORA, 2026, https://dora.dev/ai/roi/calculator/. Accessed 19 Aug. 2026.
Cui, Zheyuan Kevin, et al. “The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers.” SSRN, 22 June 2025, https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4945566. Accessed 19 Aug. 2026.
TripleTen. “Entry-Level AI Engineer Salary Calculator.” TripleTen, 2025, https://tripleten.com/tools/entry-level-ai-engineer-salary/. Accessed 19 Aug. 2026.
MIT NANDA. The GenAI Divide: State of AI in Business 2025. Cloudelligent, July 2025, https://cloudelligent.com/wp-content/uploads/2026/02/v0.1_State_of_AI_in_Business_2025_Report.pdf. Accessed 19 Aug. 2026.
Boston Consulting Group. “As AI Investments Surge, CEOs Take the Lead.” BCG AI Radar, 2026, https://www.bcg.com/publications/2026/as-ai-investments-surge-ceos-take-the-lead. Accessed 19 Aug. 2026.
Pin.com. “AI Engineer Salary Aggregation & Role Compensation Data.” Pin.com, May 2026, https://www.pin.com. Accessed 19 Aug. 2026.
Burtch Works. 2025 Artificial Intelligence & Data Science Compensation Report. Burtch Works, 2025, https://burtchworks.com/market-researchers-salary-report-2025/ai-and-data-science-2025compensationreport-introduction. Accessed 19 Aug. 2026.
Mismo. “Latin America Software Engineering Market & AI Premium Benchmarks.” Mismo, Mar. 2026, https://www.mismo.team. Accessed 19 Aug. 2026.
Robert Half. 2026 Salary Guide for Technology Professionals. Robert Half, 2026, https://www.roberthalf.com. Accessed 19 Aug. 2026.
PwC. 2025 Global AI Jobs Barometer: Wage Premiums & Skills Trends. PwC, 2025, https://www.pwc.com. Accessed 19 Aug. 2026.
Next Idea Tech. “Hire Nearshore AI Engineers Who Ship Production AI: 2026 Market Benchmarks.” Next Idea Tech, 29 May 2026, https://www.nextideatech.com/hire/ai. Accessed 19 Aug. 2026.
CEEhire. “Remote Software Developer Salaries in Central & Eastern Europe.” CEEhire, 1 May 2026, https://ceehire.com/blog/remote-software-developer-jobs-in-hungary-practical-guide-for-2026. Accessed 19 Aug. 2026.
Terminal. “Terminal Releases 2026 Remote Software Engineer Salary Insights Report.” Terminal, 9 July 2026, https://www.terminal.io/guide/remote-software-engineer-salary-insights-report. Accessed 19 Aug. 2026.
McKinsey & Company. The State of AI in 2025: Generating Value Across the Enterprise. McKinsey & Company, 2025, https://www.mckinsey.com. Accessed 19 Aug. 2026.
Anthropic. Evaluating the Impact of AI Coding Assistants on Developer Mental Models and Skill Retention. Anthropic, Jan. 2026, https://www.anthropic.com. Accessed 19 Aug. 2026.
Gartner. “Gartner Forecasts Worldwide IT Spending Growth.” Gartner Press Release, July 2026, https://www.gartner.com. Accessed 19 Aug. 2026.
Deel. State of Global Hiring 2026 Report. Deel, 2026, https://www.deel.com. Accessed 19 Aug. 2026.

Measure before you add capacity.

Start with an idea-to-production baseline. It gives your team and your engineering partner the same number to improve—and the same number to be held to.

See the AI PDLC approach