By Gowtham Raj, Director at TartLabs, writing for founders and CTOs deciding what to build in-house and what to buy.
The Short Answer
AI will not replace software development companies as a category, but it is already replacing a specific slice of what they sell, and firms that sell only that slice are in real trouble. Routine coding, boilerplate, simple internal tools and some products companies used to buy are moving in-house, handled by smaller teams with AI coding tools. Buyers now have good reason to ask whether quotes for commodity work reflect AI savings. Entry-level hiring in AI-exposed work is weaker.
What AI is not replacing is accountability. Someone still has to choose the architecture, secure the code, integrate it with older systems, turn vague requirements into something buildable, and own the result when it breaks in production. The evidence says review is still needed: more developers distrust AI output than trust it (46% against about 33%), and in a 2026 lab benchmark about 44% of AI-generated code tasks failed security checks. The companies that get replaced sell hours of routine code. The ones that survive sell outcomes. For the wider shift, see how generative AI is changing software development and delivery.
We run a development company, so read this with that interest in mind. Where the evidence goes against firms like ours, we say so.
Key Takeaways
- 32% of organisations have decided against buying at least one software product or feature because they could build it internally with agentic coding tools (McKinsey, 1,719 participants in 97 nations, 25 August 2026). These are purchase decisions already made, not a forecast.
- Computer programmer jobs are projected to fall 7% from 2025 to 2035, while software developer jobs grow 10% (US Bureau of Labor Statistics, last modified 27 August 2026). BLS expects companies to use AI to automate repetitive programming tasks.
- Employment of 22 to 25 year olds in the most AI-exposed occupations is about 19% below where it would be had it kept pace with less-exposed peers (Stanford Digital Economy Lab, revised 12 August 2026). Experienced workers show no such gap, and the authors do not claim proof of causation.
- 46% of developers distrust the accuracy of AI output and only 3.1% highly trust it (Stack Overflow Developer Survey 2025, 49,000+ respondents in 177 countries). The 2026 results are not yet published.
- Gartner predicts AI coding costs will overtake the average developer's salary by 2028 (Gartner, 24 June 2026). AI is not free labour.
- IT services spending is forecast to grow 5.3% in 2026 against 15.5% for software (Gartner, 27 July 2026). Budgets are growing, but IT services is growing at about a third of the rate of software.
What AI Is Actually Replacing in Software Work
The clearest evidence of replacement is in what organisations decide to buy. McKinsey's The state of AI in 2026: On the road to ROI, published 25 August 2026 from an online survey of 1,719 participants in 97 nations fielded between 4 May and 8 June 2026, puts it directly: "Nearly a third of respondents (32 percent) report that their organizations have decided against buying one or more software products or features because they could be built internally with agentic coding tools."
Read that as a buyer. A third of organisations have already skipped at least one purchase. McKinsey's wording covers software products and features, so some of those decisions replaced SaaS subscriptions rather than agency work. But the build-versus-buy calculation, which we set out in custom software versus generic software, is clearly being redone by more organisations.
The same survey shows how early the shift still is: about two in ten organisations are scaling software coding agents, rising to 31% at larger enterprises. It also reports that 37% attribute at least some EBIT impact to AI use, about the same share as last year. The two figures measure different things, so they are not a like-for-like comparison.
At Google, the share of AI-written code has jumped. At Google Cloud Next on 22 April 2026, Sundar Pichai said: "75% of all new code at Google is now AI-generated and approved by engineers, up from 50% last fall." It is self-reported by Google, and it counts code that engineers approved. It measures who drafts the code, not who answers for it.
Public employment projections point the same way at the task level. The BLS Occupational Outlook Handbook for computer programmers, last modified 27 August 2026, projects a 7% decline from 2025 to 2035, from 110,800 jobs to 102,700, and explains why: "many companies are expected to leverage technologies, including artificial intelligence (AI), to automate repetitive programming tasks." In the BLS classification, programmers write and test code largely to designs someone else has specified. In our view that is the slice of work AI tools handle best, and it overlaps heavily with what some agencies sell by the hour.
Put together, the work being absorbed looks like this:
- Boilerplate and scaffolding: project setup, data models, standard API endpoints, form validation.
- Simple internal tools: admin panels, reporting dashboards, CRUD apps and scripts that an internal team with AI tools can now produce itself.
- Well-documented integrations: connecting two modern systems that both publish clean APIs.
- Some bought software: small features and niche tools that once justified a subscription or a vendor.
- First drafts of tests and documentation, which still need review but no longer need a dedicated person to start.
If a development company's revenue sits mostly in that list, it is the most exposed to AI. McKinsey's figure shows buyers are already changing what they purchase.
The Hiring Data: Fewer Junior Seats, Not Fewer Developers
The labour data does not show software development disappearing; it shows the work changing shape, with the entry level taking the hit. The BLS Occupational Outlook Handbook for software developers, last modified 27 August 2026, projects software developer employment to grow 10% from 2025 to 2035, from 1,717,800 jobs in 2025, with about 106,100 openings a year across software developers, quality assurance analysts and testers. QA analysts and testers are projected to grow 6%. Median pay for software developers was $135,980 in May 2025. These projections were revised after AI coding tools went mainstream, so they are not a pre-AI artefact.
Job postings tell a colder story. Indeed Hiring Lab's index of US software development job postings, published through FRED as series IHLIDXUSTPSOFTDEVE, stood at 77.32 on 18 September 2026, where 1 February 2020 equals 100. Postings remain about 23% below their February 2020 baseline. Postings are not jobs, though, and the index also reflects interest rates, the unwinding of the post-pandemic hiring boom and shifting budgets. It cannot be pinned on AI alone.
The most careful attempt to isolate AI's effect comes from Stanford. In Canaries in the Coal Mine?, Brynjolfsson, Chandar and Chen of the Stanford Digital Economy Lab (revised 12 August 2026, using ADP payroll data through June 2026) find that employment of workers aged 22 to 25 in the most AI-exposed occupations is about 19% below where it would be had it kept pace with less-exposed peers. Experienced workers in the same occupations show no such gap. The effect runs mainly through reduced hiring rather than layoffs. The authors caution that this is not proof of causation, and they find no evidence of economy-wide displacement.
The tasks that used to train junior engineers, such as boilerplate, small tickets and simple fixes, are the tasks AI tools handle most readily. That is our reading rather than a measured result, but it suggests a vendor built on a wide pyramid of junior engineers billed by the hour has more to explain than one that staffs smaller, more senior teams. If you are weighing that against building your own team, our comparison of dedicated developers versus an in-house team covers the control and cost trade-offs.
Does AI Make Developers Faster? The Evidence Is Still Unsettled
No one can yet tell you with confidence how much faster AI makes experienced developers, and anyone quoting a precise multiplier is overselling. The best controlled evidence has moved in both directions within a year.
In July 2025, METR published a randomised controlled trial finding that experienced open-source developers took 19% longer to complete tasks when using early-2025 AI tools. On 24 February 2026, METR published an update covering 57 developers and more than 800 tasks. It found 18% less task time for returning developers and 4% less for new recruits, as point estimates with wide confidence intervals that include no change. METR itself called this "very weak evidence", partly because developers increasingly refused to work without AI, which skews who takes part in a study that asks some of them to go without. Neither result is settled, and neither should be quoted as if it were.
Google's 2025 DORA report, a survey of about 5,000 technology professionals, found that 90% use AI at work and over 80% say it raised their productivity, while 30% report little or no trust in AI-generated code. Its central finding matters most to a buyer: AI is an "amplifier" of existing strengths and weaknesses. A team with good tests, continuous integration, disciplined code review and clear requirements gets faster. A team without them produces more code, faster, with the same defects at higher volume. That applies equally to your internal team and to any vendor you hire. When a development company tells you AI makes it faster, the follow-up question is what it is amplifying.
Why AI Output Still Needs an Accountable Owner
AI can write code, but it cannot be accountable for it, and the trust and security data both say someone has to be. Developers, the people closest to the output, are the most sceptical of it.
The Stack Overflow Developer Survey 2025, with more than 49,000 respondents across 177 countries, found 84% use or plan to use AI tools. But 46% distrust the accuracy of AI output, against about 33% who trust it, and only 3.1% highly trust it. 66% cite AI solutions that are "almost right, but not quite" as a frustration, and 45.2% say debugging AI-generated code is more time-consuming. The 2026 survey results have not been published yet, so these remain the latest figures.
Security testing points the same way. Veracode's 2026 GenAI Code Security Report, published 28 July 2026, found AI-generated code averaged a 56% security pass rate across the tested tasks, meaning about 44% introduced a vulnerability. The best model reached 68%. On specific weakness classes the results were worse: cross-site scripting checks passed only 15% of the time, and log injection 12%. These are lab benchmark tasks rather than production systems, so they do not tell you the defect rate in any given codebase. They do tell you that unreviewed AI output is not safe to ship by default.
This is the part of a development company's work that AI is not replacing, and it is where buyers should focus when they decide what to pay for:
- Architecture. Choosing boundaries, data models and trade-offs that will hold for years. AI can propose options; someone has to choose and live with the consequences.
- Security. Threat modelling, secure defaults, dependency hygiene and review of AI output against weakness classes like the ones Veracode tested.
- Integration. Connecting to ERPs, payment rails, identity systems and old databases whose behaviour is documented nowhere. This is where most real projects spend their time, as our guide to integrating AI into business software shows.
- Requirements. Turning "we need a portal" into a scoped, testable specification. AI accelerates building the wrong thing as readily as the right one.
- Production ownership. Monitoring, incident response, upgrades and the person who picks up the call. Legacy systems in particular need that continuity, which is why legacy software modernisation remains a people-heavy discipline.
AI Is Not Free: The Cost Curve Buyers Miss
The in-house AI build is cheaper than it used to be, but it is not free, and the cost is rising in a way most business cases ignore. On 24 June 2026, Gartner predicted: "By 2028, AI coding costs will overtake the average developer's salary due to rising large language model (LLM) token consumption and the shift to consumption-based licensing models."
That prediction changes the build-in-house arithmetic. A flat per-seat licence made AI coding tools look like a rounding error. Agentic tools that run long tasks, retry, read whole repositories and call models repeatedly consume tokens at a rate that scales with ambition. If your plan to replace an agency rests on "the tools cost a few dollars per user a month", check whether that is still the pricing model you will be on in two years.
Gartner's recommendation is practical and worth adopting whether you build internally or hire a partner. It suggests classifying tasks into three groups: developer-led, developer-with-agent, and fully agent-led. Fully agent-led work should be routine and cheap to verify. Developer-led work is where judgement, security and architecture live. The middle category is where most of the productivity gain and most of the risk sit together.
Some of the internal builds behind McKinsey's 32% figure will prove to be good decisions. Some will become unowned internal tools that nobody maintains, with a token bill attached. Both outcomes are plausible, and the survey data cannot yet tell us the ratio.
What AI Means for Software Development Pricing
Buyers should ask whether pricing reflects AI savings, and prefer pricing tied to outcomes where the work can be defined. The spending data shows where the money is going.
Gartner's 27 July 2026 forecast puts worldwide IT spending at $6.37 trillion in 2026, up 14.2%. Within that, software grows 15.5% to $1,468 billion, IT services grows 5.3% to $1,570 billion, and data center systems grow 62.5% to $822 billion, driven by AI infrastructure. Budgets are expanding, but IT services is growing at about a third of the rate of software, and the fastest growth is in AI infrastructure and software.
One large services firm gives a data point. In its Q1 FY27 results, filed with the SEC on Form 6-K on 23 July 2026, Infosys gave FY27 revenue growth guidance of "1.5%-3.0% in constant currency". That is modest, though the two figures are not like for like: one is a single company's constant-currency revenue guidance and the other is a worldwide spending forecast. Many factors feed into guidance, and we do not claim AI is the cause.
India's industry body describes the same transition. NASSCOM's Strategic Review 2026, published in February 2026, expects India's technology sector to cross $315 billion in FY26, while direct employment of about 6 million grew only 2.3%, a net addition of 135,000. NASSCOM frames this as a shift "from scale-led growth to value and innovation" and toward outcome-based, risk-sharing contracts. Headcount growth of 2.3% sits alongside a sector review that describes a move away from scale-led growth.
What this means for your next contract:
- Ask whether commodity work is priced with AI in mind. If a quote for boilerplate-heavy work looks the same as it did three years ago, ask why.
- Prefer fixed-scope or outcome-based pricing where the work is definable. It moves estimation risk, and AI's efficiency gains, to the party doing the work, and it rewards the vendor for finishing rather than for hours.
- Keep time-and-materials for genuinely uncertain work, such as discovery, research spikes and legacy archaeology, where no one can price the outcome honestly.
- Ask how AI savings are shared. A vendor using AI tools heavily while billing the same hours is keeping the entire gain.
For the wider Indian vendor market, our overview of the offshore software development company model in India covers how engagement structures are changing.
What to Keep Buying From a Dev Partner and What to Do In-House
Do the work in-house with AI tools when it is routine, cheap to verify and low-risk if wrong; buy from a partner when it needs accountable judgement, specialist depth or production ownership you do not have. The table applies that rule.
| Work | In-house with AI tools | Buy from a partner | Why |
|---|---|---|---|
| Internal admin panels, dashboards, scripts | Yes | Rarely | Routine, low blast radius, easy to check |
| Clickable prototypes and demos | Yes | Optional | Throwaway by design; speed matters more than structure |
| Customer-facing MVP meant to carry real users | If you have senior engineers | Often | Early architecture decisions persist long after launch |
| Architecture for a new platform | Only with senior internal ownership | Yes | Judgement-heavy and expensive to reverse |
| Security review and hardening | Basic scanning | Yes, or an independent reviewer | AI output fails common weakness checks in benchmark testing |
| Integration with ERPs, payments, legacy systems | Simple, well-documented APIs | Yes, for the rest | Undocumented behaviour is where projects slip |
| Legacy modernisation | Code reading and test generation | Yes | Needs continuity, migration planning and rollback |
| Regulated or compliance-bound features | No | Yes, with audit trails | Accountability has to be explicit and documented |
| Maintenance of well-tested code | Yes | Optional | Tests make AI changes cheap to verify |
| Production on-call and incident ownership | If you staff it | Yes, if you do not | Someone has to answer at 2 a.m. |
The dividing line is the cost of being wrong and whether you have someone senior enough to catch it. Prototypes are an obvious in-house job now, which costs agencies real work. A prototype and a product are still different things. Our guides to AI MVP development and the broader MVP development process cover where that line usually falls.
Our mapping of the table onto Gartner's three categories, not Gartner's own advice, gives a staffing plan. Fully agent-led work belongs in-house. Developer-with-agent work can sit either side, depending on whether you have the reviewers. Developer-led work, where architecture, security and integration live, is where a partner earns its fee or where you need senior hires. If you are deciding whether to build AI features at all rather than buy them, custom AI development versus off-the-shelf AI works through that choice.
Questions to Ask a Development Company About How It Uses AI
Any development company you hire in 2026 uses AI tools; the questions that matter are how it prices, reviews and secures that use. A vendor that cannot answer these clearly is either not using AI deliberately or not telling you how.
On pricing:
- Which parts of this estimate assume AI assistance, and how much time does that save?
- How are AI efficiency gains reflected in the price: lower hours, a fixed price, or a shared-savings clause?
- Who pays for AI tool licences and token consumption, and is that a pass-through with a cap?
- If we move to outcome-based pricing, what are the acceptance criteria?
On review and quality:
- What share of delivered code is AI-generated, and what review does every AI-generated change go through?
- Who is the named senior engineer accountable for architecture and for merge approval?
- Which automated tests, static analysis and security scans run before code reaches us?
- How do you handle "almost right" output, the kind 66% of developers in Stack Overflow's survey complain about?
On IP, data and security:
- Which AI tools and models touch our code, and under which data-retention and training terms?
- Is our source code, data or documentation ever sent to a model that could train on it?
- Does IP assignment cover AI-assisted output explicitly, and what do you warrant about third-party code in it?
- How do you check AI-generated code for the weakness classes that fail most often in benchmarks, such as cross-site scripting and log injection?
Ask the same questions of every vendor on your shortlist, including us. If you are building that shortlist from scratch, our list of software development companies in India is one starting point, and these questions will separate firms using AI to deliver better outcomes from firms using it only to protect margins.
The Stats That Do Not Survive Checking
Several numbers in this debate get repeated without their date or caveat. Each correction below is against the publisher's own page or a named outside analysis.
"The Stack Overflow 2026 survey shows 84% use AI but only 3% highly trust it." Those are 2025 figures relabelled. The 2026 survey opened on 23 June 2026, and its results have not been published. The 2025 results are the latest available, and they should carry the 2025 label.
"METR proved AI makes developers 19% slower." The July 2025 trial used early-2025 tools and experienced open-source developers. METR did not retract it, but its 24 February 2026 update reported point estimates of less task time with confidence intervals that include no change, and called the data "very weak evidence". Neither study proves anything settled in either direction.
"Stanford found AI wiped out 20% of junior developer jobs." The original figure was a decline of about 20% from a late-2022 peak in employment of software developers aged 22 to 25. The authors do not claim it is entirely driven by AI, the effect is reduced hiring rather than layoffs, and the current headline finding is a relative gap of about 19% against less-exposed peers.
Older, lower figures for how much of Google's code is written by AI. Google now says 75% of new code is AI-generated and approved by engineers, up from 50% last fall. That figure is self-reported and depends on human approval, so it supports the case for accountable engineers as much as the case against them.
The underlying trends are real: entry-level hiring in exposed work is weaker, AI writes a growing share of code, and productivity evidence is mixed. Quote the figure the source published, with its date and caveats.
When You Do Not Need a Development Company
Sometimes you should not hire one, and we would rather say so than sell work that AI tools and your own team can now handle.
When the work is a routine internal tool. If it is a dashboard, an admin screen or a workflow script used by your own staff, with low consequences if it breaks, an internal developer with AI tools is often the better choice than a vendor engagement.
When you have senior engineers with spare review capacity. AI output is cheap; review is the constraint. If you already have people who can own architecture and review AI changes, much of what you would buy from a partner is available in-house.
When the need is a prototype to test an idea. A clickable demo or a throwaway proof of concept is now something one capable person can build with AI tools. Paying an agency to build it the traditional way is harder to justify.
When an off-the-shelf product already fits. Neither AI nor a custom build beats a mature product that already does the job. Custom work only earns its cost when the software is a differentiator or the fit is genuinely poor.
The situations where a partner still earns its fee are the opposite of these: high consequences if wrong, specialist depth you lack, integration with systems you cannot easily see into, and production ownership you cannot staff.
The Bottom Line
AI is replacing part of what software development companies sell. A third of organisations in McKinsey's 2026 survey have already skipped a software purchase to build it with coding agents, and BLS projects computer programmer jobs to shrink while developer jobs grow. The firms most exposed are the ones whose value was hours of routine code.
The accountable part of the work remains. More developers distrust AI output than trust it, about 44% of AI-generated code tasks failed Veracode's security checks, and Gartner expects AI coding costs to overtake a developer's salary by 2028. Architecture, security, integration, requirements and production ownership still need a person or a partner who answers for the result.
The practical decision is which work to keep in-house with AI tools, which to buy as an outcome, and how to hold any vendor to account for the way it uses AI. Build the routine work yourself, buy judgement and ownership where you lack them, and ask that AI savings show up in the price.
Weighing what to build in-house against what to hand to a partner? If a candid second opinion on your roadmap would help, including which parts you no longer need to outsource, get in touch.
