AI-Assisted Software Delivery: A Buyer's Guide
AI-assisted software delivery means engineers use AI coding tools inside an ordinary engineering process. Before you pay a vendor for it, test the claim on your own codebase: measure lead time, review time and defects, and ask who reviews AI-drafted changes, where your code goes and who pays for rework. This guide is written by StepTo, a 15-20 engineer nearshore team in Belgrade, makes no AI speed-up claim for StepTo, and shows how to test the claim from any vendor.
Reviewed by Igor Gazivoda, Founder & CEO of StepTo · Updated
What Should You Ask Before Paying for AI-Assisted Delivery?
Ask for evidence you can reproduce on your own repository: a baseline, a measurement period, the kind of code it was measured on, and the contract clause that depends on it. Then ask who reviews AI-drafted changes, which tools see your code and who pays for defects found after merge. StepTo is a 15-20 engineer nearshore team in Belgrade working on Central European Time, and the ten questions below apply to StepTo as much as to any other vendor.
The market is loud on this topic. Several large vendors now sell AI pods, AI-native delivery frameworks or agentic engineering programmes, and most state a speed-up. The sections below show how to measure the effect yourself, what independent research has found so far, how to read the vendor figures side by side, and which clauses to put in the contract.
What Should You Measure on Your Own Codebase?
Measure the same team on the same kind of work before and after the tool arrives, using delivery outcomes rather than lines of code or the number of AI suggestions accepted. DORA defines five software delivery performance metrics that make a sound starting point, and three further measures cover what an AI-assisted workflow is most likely to shift.
The five metrics below are summarised from the DORA metrics guide, read on 2026-11-20. DORA groups the first three as throughput and the last two as instability.
Change lead time
The time from a change being committed to version control until it is deployed in production.
Deployment frequency
How often deployments occur over a period, or the gap between them.
Failed deployment recovery time
How long it takes to recover from a deployment that fails and needs immediate intervention.
Change fail rate
The share of deployments that need immediate intervention afterward, such as a rollback or hotfix.
Deployment rework rate
The share of deployments that are unplanned because of a production incident.
Three Measures DORA Does Not Define
These are our suggestions, not DORA metrics. They exist to catch a saving that moves from one stage to another instead of disappearing.
Review time per pull request
How long a change waits for a reviewer and how long the review takes. If drafting gets faster and review gets slower, the saving moved rather than appeared.
Defects found after merge
Bugs that reach staging or production from changes that passed review. Count them per period, not per feature, so a busy month does not hide them.
Rework
The share of merged work that is reopened or rewritten within a fixed window that you choose in advance, for example 30 days.
What Does Independent Research Say So Far?
Careful measurement has produced a changing and uncertain answer on speed, and a steady and unflattering one on security. Two sources read on 2026-11-20 show why a vendor figure needs a baseline, a date and an interval before it earns trust.
METR: randomized trials on experienced developers
In the July 2025 study, 16 experienced open-source developers worked on 246 real issues in repositories they knew well. When they were allowed to use AI tools they took 19% longer. They had expected a 24% speed-up and still believed AI had sped them up by 20% afterwards. METR now marks those results as out of date.
In its February 2026 update (57 developers, 143 repositories, 800+ tasks), METR estimates task times 18% shorter with AI for the 10 returning developers, with a confidence interval from 38% shorter to 9% longer, and 4% shorter for the 47 new ones, with an interval from 15% shorter to 9% longer. METR says selection effects make its data "only very weak evidence" for the size of the increase and that its estimate is likely a lower bound.
Veracode: security of AI-generated code
In the Spring 2026 GenAI Code Security update, models were given 80 coding tasks in four languages with no security-specific guidance. Across all models and tasks, only 55% of generation tasks resulted in secure code, so in 45% of cases the model introduced a known security flaw.
The test measures models writing code out of the box, not a team with review and scanning in place. Read the 45% as a reason to insist on a review gate and a pipeline scan, not as a forecast for your repository.
The pattern in the METR work: a believed gain that differed from the measured one, intervals that include zero, and an answer that moved within a year. The Veracode figure is about security rather than speed, and it is why a review gate and a scan matter. For the delivery side of the same story, see the StepTo blog posts on AI coding tools and release speed and on comprehension debt.
How Should You Read Vendor Productivity Multipliers?
Read a multiplier as a claim about a baseline, not as a fact about your project. Ask what it is measured against, whether it was measured or modeled, on what kind of work, and which page of the vendor's site the contract will quote. The table sets out what nine pages from five vendors report, as of 2026-11-20.
| Vendor | Page | Reports | Basis stated |
|---|---|---|---|
| Software Mind | /ai-pods/ | "2-3x more output" and "~50% reduction in idea-to-production lead time" | Vendor statement. No method found on the page. |
| Software Mind | /ai-information/ | "up to 10x boost" | Vendor statement. No method found on the page. |
| Software Mind | /who-we-are/our-company/ | "5-10x" | Vendor statement. No method found on the page. |
| STX Next | /services/ai-native-sdlc | Tier 2: 20 to 50% faster to market and 30% cheaper. Tier 3: 80%+ faster and 60% cheaper. Both against a traditional team of 6 to 8 engineers. | Labelled "Modeled estimates from STX Next's delivery simulator". |
| STX Next | /about-us | "2x-5x for brownfield, 10x+ for greenfield", a 2-10x range | Vendor statement. |
| STX Next | /our-partners/anthropic | "up to 60% cost reduction and up to 5x-10x faster" | Vendor statement. |
| Endava | /dava-flow | "typically reduces time-to-market by up to 40 per cent" | Vendor statement. |
| Ciklum | /services/product-engineering-services/ | "AI accelerates delivery by up to 40%" | Vendor statement. |
| Andela | /learn/agentic-engineering | "+30.5% Modeled Productivity Lift" | Modeled. Sourced on the page to Andela's own study of 41 engineers over 2 weeks. |
Worked example. Software Mind reports 2-3x more output on /ai-pods/, up to 10x on /ai-information/ and 5-10x on /who-we-are/our-company/. The top of the first range sits below the bottom of the third. The pages may describe different work, and that is the point: a buyer needs to know which figure a contract would rely on. STX Next gives three different kinds of figure on three pages, from tier estimates to a 2-10x range to up to 5x-10x, and at least labels its tier figures as modeled estimates from its own simulator, which tells the reader what sort of number it is. A bare multiplier does not.
What is the baseline?
Faster than what: the same team last year, a hypothetical team of 6 to 8 engineers, or an industry average? A multiplier without a baseline cannot be checked.
Measured or modeled?
A measurement happened on real work. A model is an estimate from assumptions. Both can be useful, but they are different kinds of number.
On whose work?
Greenfield code and a ten-year-old monolith behave differently. A range that spans both, such as 2x to 10x, tells you little about yours.
Which page will the contract quote?
If a vendor's own pages give different figures, ask which one the statement of work refers to, and what happens to the price if you do not see it.
Where Do Other Vendors Simply Win?
On certifications, published AI programme detail, scale and named references, the larger vendors above are ahead of a 15-20 engineer team such as StepTo, and a buyer whose requirements include those should shortlist them.
Certifications
Endava reports SOC 2 Type II and ISO 27001 (covering approximately 60% of its legal entities as of June 2025) on /who-we-are/certifications, and HTEC reports ISO 27001 and SOC 2 Type II on /quality-first/, both as of 2026-11-20. If your procurement requires a certified supplier, those vendors meet that requirement and StepTo is not the right shortlist entry.
Programme detail
STX Next publishes pod tiers, a stated modeling basis and governance controls on /services/ai-native-sdlc, and Andela publishes an assessment on /assess/agentic-engineering plus three training levels and a calculator on /learn/agentic-engineering, as of 2026-11-20. If you want a vendor-run AI adoption or training programme, they publish far more than StepTo does.
Scale and references
Software Mind reports 1600+ experts on /ai-information/ and Ciklum reports 3,000+ engineers on /about-ciklum/, both as of 2026-11-20, against StepTo's 15-20 engineers. Andela publishes a named case study at /case-study-github. For a large, multi-country programme with named references, they win on scale.
Which Contract Clauses Matter for AI-Assisted Delivery?
Eight clauses cover most of the risk: tool disclosure, data handling, IP ownership, a review gate, tests and scans, rework, reporting and exit. The right-hand column shows which of them StepTo's published practices already cover, and which you should ask any vendor, StepTo included, to write down. This is buyer guidance, not legal advice.
| Clause | What to require | In StepTo's published practices |
|---|---|---|
| Tool disclosure | A written list of the AI tools that touch your code or data, and notice before the list changes. | Not one of the eight published practices. Ask every vendor, StepTo included, to put it in writing before you sign. |
| Data handling | What code and data leave your environment, where they go, and whether the tool provider may train on them. Andela reports "No model training" on client data on /deploy/managed-ai-solutions as of 2026-11-20, which is the kind of line to ask for. | Not one of the eight published practices. Ask every vendor, StepTo included, to put it in writing before you sign. |
| IP ownership | An assignment that covers all work product, however it was produced, signed before the first engineer starts. | An NDA and an IP assignment are signed before any engineer starts, and all IP goes to you. |
| Review gate | A named human reviewer, who is not the author, approves every change before it is merged. | Every change is reviewed by a second engineer before it is merged. |
| Tests and scans | Automated tests run on every pull request. For AI-drafted code, also ask for a security scan in the pipeline. | A CI pipeline runs automated tests on every pull request. A security scan is not one of the eight published practices; ask for it in writing. |
| Rework | Who pays for defects found after merge, over what window, and how a defect is defined. Agree it in the statement of work, not in an email. | Not one of the eight published practices. Agree it in the statement of work. |
| Measurement and reporting | Access to your own repository metrics and a regular report, so you can compute lead time and review time yourself. | You get a demo or a written status report every week. |
| Exit | Code, tickets and documentation live in your accounts from day one, so leaving never depends on a handover negotiation. | Engineers work in your repository and your tools from day one, and you own the code at all times. |
The statement of work clause on the engineering practices page restates the practices in contract language for any vendor. The vendor scorecard and the due diligence answers cover the rest of a procurement review.
How Do StepTo's Eight Engineering Practices Fit?
StepTo applies eight engineering practices on every engagement, and each one comes with a check you can run yourself in the first weeks. The review gate and the test run apply to every change, whoever or whatever drafted it, and you can check both on any merged pull request. The full list, with how to verify each, is on the engineering practices page.
Code and delivery
Security and continuity
For embedded engineers working inside your systems on AI projects, see forward deployed engineers. For building AI features into a product, see AI and automation solutions.
What Can You Compare on Price Today?
StepTo publishes its prices, so you can set them against any AI-assisted offer you are quoted. Staff augmentation starts from $4,500 per developer per month, with a 3-month minimum and weekly billing. A dedicated team of three starts from $13,500 per month, billed monthly, and projects and MVPs start from $9,000.
| Seniority | Hourly rate |
|---|---|
| Junior | $25-35/hr |
| Mid-level | $35-45/hr |
| Senior | $50-65/hr |
| Lead | $65-85/hr |
A pricing page was not found in the STX Next sitemap (570 URLs) as of 2026-11-20, nor in the Ciklum page sitemap as of the same date. StepTo clients typically see 36-51% lower costs than hiring locally in Western Europe or the US, a difference between regions and not a measured effect of any tool. A rate card prices engineers' time and a quoted multiplier prices a result, so ask any vendor that prices a result to show the baseline it priced against. Full terms are on the pricing page.
What Ten Questions Should You Put to Any Vendor?
Put these ten questions to every vendor before you sign, and ask for the answers in writing. Vague answers are an answer too.
- 1Which AI tools touch our code and data, and who approved each one?
- 2Do the tool providers store our prompts or code, and may they train on them?
- 3Who reviews AI-drafted changes, and can we see the reviewer on each pull request?
- 4Does the pipeline run tests and a security scan on every pull request?
- 5What baseline does your productivity figure compare against, and was it measured or modeled?
- 6On what kind of codebase was it measured, and with what confidence interval?
- 7Which page or document will the contract quote for that figure?
- 8Who pays for defects found after merge, and over what window?
- 9Will we get access to lead time, review time and defect data for our own repository?
- 10What changes in the price if we do not see the improvement you described?
How Do You Start With StepTo?
It starts with a 15-minute intro call. If there is a fit, a 60-minute discovery call follows. Engineers can typically start within one to two weeks of first contact; a full team usually takes two to four weeks. An NDA and IP assignment terms are signed before any work begins. Bring the ten questions above to the discovery call.
StepTo replaces an engineer free of charge if the fit is wrong within 30 days. After the 3-month minimum, scaling down takes 30 days' notice. Staffing options are on the staff augmentation and dedicated team pages.
What Would an Engineer Cost?
Use our calculator to estimate a cost range for your team size and seniority, based on StepTo's published rates.
Free · No signup
What Do Buyers Ask About AI-Assisted Delivery?
Does StepTo claim that AI coding tools make delivery faster?
This guide makes no speed-up claim for StepTo and gives no AI productivity figure. StepTo clients typically see 36-51% lower costs than hiring locally in Western Europe or the US, a difference between regions and not a measured effect of any tool. To learn what a tool does for your team, measure lead time, review time and defects on your own codebase before and after.
Who owns the code StepTo engineers write?
You do. StepTo engineers work in your repository and your tools from day one, and an IP assignment is signed before any engineer starts, so all IP goes to you.
How can I check StepTo's review and testing before I commit?
Because the work happens in your repository, you can open any merged pull request and see an approving review from someone other than the author, plus the automated test run next to it. You should also have received two weekly demos or reports by the end of the first two weeks.
What does StepTo cost?
StepTo's rate card runs $25-85/hr: junior $25-35, mid $35-45, senior $50-65 and lead $65-85. Staff augmentation starts from $4,500 per developer per month, with a 3-month minimum and weekly billing. A dedicated team of three starts from $13,500 per month, billed monthly. Projects and MVPs start from $9,000.
How does an engagement with StepTo start?
It starts with a 15-minute intro call, and if there is a fit, a 60-minute discovery call follows. Engineers can typically start within one to two weeks of first contact; a full team usually takes two to four weeks. An NDA and IP assignment terms are signed before any work begins, and StepTo replaces an engineer free of charge if the fit is wrong within 30 days.
Or reach us directly
Office hours
Sent on a weekend: we reply on the next business day.
Contracting party: StepTo d.o.o. Company details are on the Imprint page.
Send us a message
StepTo's Belgrade team replies within one business day.
It starts with a 15-minute intro call. If there is a fit, a 60-minute discovery call follows. Engineers can typically start within one to two weeks of first contact; a full team usually takes two to four weeks.
