Skip to main content
Back to News & UpdatesIndustry Insight

AI proof of concept: what it costs, how long it takes, and how to get it into production

|13 min read
An illustrated robot holding a glowing laboratory test flask, beside the headline AI Proof of Concept to Production. Test one use case on real data against an agreed pass mark, then plan the move to production.

An AI proof of concept tests one use case on real data against an agreed pass mark. Timeline, cost arithmetic and a checklist for getting to production.

Key takeaways

  • An AI proof of concept tests one use case on real data against a pass mark agreed before the build, and by our planning estimates takes two to six weeks.
  • Judge a proof of concept on four numbers: the score on held-out test cases, cost per completed task, latency at the slow end, and the rate and severity of failures.
  • One AI engineer for four weeks is 160 × $45 = $7,200 at the hourly rate or $6,100 on the 160-hour monthly plan, and US Eastern overlap adds 15%.
  • In a 2025 S&P Global Market Intelligence survey, the average company abandoned 46% of its AI proofs of concept before production.
  • IDC found that for every 33 AI proofs of concept a company launched, only four reached production (March 2025).
  • Before calling a proof of concept done, name the owner, the monthly run budget, the monitoring plan and the security review.

An AI proof of concept is a short build that tests whether one AI use case works on your own data, against a pass mark agreed before work starts; by our planning estimates, it takes two to six weeks. It answers one question: can AI do this task well enough, cheaply enough and fast enough to be worth building properly?

Most of the cost is engineering time. One AI engineer for four weeks is 160 hours, which the worked example below prices at $6,100 to $7,200 at our published rates. The harder step comes after: in a 2025 S&P Global Market Intelligence survey, the average company abandoned 46% of its AI proofs of concept before they reached production (Fortune). All figures here were checked on 17 November 2026.

AI proof of concept vs pilot vs MVP

The terms are used loosely, but they answer different questions. A proof of concept (PoC) asks whether the AI can do the task. A pilot asks whether it works for real users in a real workflow. An MVP (minimum viable product) asks whether customers will use it and pay for it.

Microsoft's Cloud Adoption Framework says a proof of concept helps reduce "implementation risk by validating technical feasibility and business value before full-scale development" (Microsoft). It also advises starting "with internal, non-customer-facing projects to limit risk".

Swipe the table sideways to see every column.

StageQuestion it answersWho uses itWhat happens to the code
Proof of conceptCan AI do this task well enough on our data?The project team and a few expert reviewersOften reworked; the test set and findings carry forward
PilotDoes it work in the real workflow, at real volume?A limited group of real usersHardened and kept
MVPWill customers use it and pay for it?Early customersBecomes the first product version
ProductionCan we run it safely every day?Everyone in scopeMonitored, maintained and retested

UK readers may know the government version. The GOV.UK Service Manual says alpha prototypes should be "just complex enough to let you test different ideas, not production quality code", and teams should "expect to throw away any code" (GOV.UK). Private beta then means "inviting a limited number of people to use your service" (GOV.UK).

A staffing trial is not a PoC either: a trial tests the engineer, a PoC tests the use case. Our guide to running a two-week trial covers the first.

What a good AI proof of concept tests

A good AI proof of concept tests six things, with targets written down before the build. Without them, it is a demo.

  • One use case: one task, one user group and one or two systems.
  • Real data: the documents, tickets or records the system will really see, messy ones included.
  • An evaluation set with a pass mark: real inputs with agreed answers, some held back for the final run, and a pass/fail score signed off before the first prompt is written.
  • Cost per task: model fees per completed task, retries included.
  • Latency: how long users wait, at the slow end as well as on average.
  • Failure modes: what goes wrong, how often and how badly.

Write the pass mark before the build

Anthropic's guidance says good success criteria are "specific", "measurable", "achievable" and "relevant" (Anthropic). Instead of "classify sentiments well", its example asks for an F1 score of at least 0.85 on a held-out test set, with 95% of responses under 200 milliseconds.

OpenAI's guidance names the opposite as an anti-pattern: "vibe-based evals", meaning "it seems like it's working" as an evaluation strategy (OpenAI).

For example, a ticket-routing pass mark might read: on 300 held-out real tickets, at least 90% reach the right queue, none leak personal data, 95% of responses arrive within three seconds, and model fees stay under $0.02 a ticket. You set the numbers; what matters is that they exist before the build.

A public example of a threshold at work

The UK government's AI Playbook (February 2025) describes GOV.UK Chat, a prototype that answered questions from GOV.UK content using retrieval-augmented generation (RAG). The project was managed through "a series of phased experiments, each lasting a couple of weeks". With "an answer accuracy of 80%", the team concluded that "accuracy gains would be required" before deploying it, and planned a limited public pilot only "provided that accuracy thresholds could be met". One number, one clear decision.

Work out the cost per task

Model fees are billed per million tokens. As of 17 November 2026, Anthropic lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, and Claude Haiku 5.5 at $0.10 and $0.50 (Anthropic). OpenAI lists gpt-6.1-sol at the same $2 and $10, and gpt-6-luna at $0.10 and $0.50 (OpenAI).

The example below assumes 3,000 input and 500 output tokens per task; your PoC should measure the real figures.

Swipe the table sideways to see every column.

Price tier (per million tokens)Input costOutput costPer task10,000 tasks a month
Mid-priced: $2 in, $10 out3,000 × $2 ÷ 1M = $0.006500 × $10 ÷ 1M = $0.005$0.011$110
Small: $0.10 in, $0.50 out3,000 × $0.10 ÷ 1M = $0.0003500 × $0.50 ÷ 1M = $0.00025$0.00055$5.50

An agent that calls the model five times per task costs roughly five times as much, often more as its context grows. Anthropic notes that agentic systems "often trade latency and cost for better task performance" (Anthropic). OpenAI's Batch prices for the two models above are half the standard rates, which suits work that does not need an instant answer.

Measure latency the way users feel it

Anthropic defines time to first token as the time from sending the prompt to the first token of the response, which matters especially when answers are streamed (Anthropic). Report it with total response time, for the slowest 5% as well as the average. Anthropic also advises getting a prompt to work first, "and then try latency reduction strategies afterward". In a PoC, reach the pass mark on a capable model, then test whether a smaller, faster one still passes.

Look for the failure modes

Average scores hide the cases that hurt. Anthropic's edge cases to test include missing or irrelevant input, overly long input, harmful input, and ambiguous cases where "even humans would find it hard to reach an assessment consensus".

The OWASP Top 10 for LLM applications (2025) adds security failures: prompt injection, sensitive information disclosure, excessive agency, misinformation and unbounded consumption. A good PoC report lists failures by type and severity, not just a pass rate.

How long an AI proof of concept takes

By our planning estimates, two to six weeks of build time: about four for one task on one data source, closer to six with a live integration or a second source. Calendar time is often longer, because access and approvals come first.

WhenWorkOutput
Before week 1Data and system access, data protection sign-off, test set collected, pass mark agreedA one-page brief signed by the owner
Week 1Simplest working version run against the test setBaseline scores
Week 2Improve retrieval, prompts and model choice; review every failureScores for each change
Week 3Cost per task, latency at the slow end, adversarial inputsCost, latency and failure figures
Week 4Final run on held-out cases, write-up, decision meetingGo, change scope, or stop

This is a planning estimate, not a quote. For reference, GOV.UK alphas "tend to last between 6 and 8 weeks" (GOV.UK). MIT NANDA's 2025 research found that mid-market firms moved from pilot to full implementation in about 90 days, against nine months or longer for large enterprises (Virtualization Review).

The usual delays are waits for API keys, a redacted data extract or a security reviewer, so start that paperwork during scoping. For the whole path from discovery to a live agent, see our AI agent timeline guide.

What an AI proof of concept costs: worked arithmetic

Mostly engineering time. The table multiplies Webify.AI's published rates by the hours in each scenario. It illustrates the arithmetic only; it is not a package or a quote.

Swipe the table sideways to see every column.

IllustrationHoursArithmeticCost
One AI engineer, 4 weeks, hourly160160 × $45$7,200
One AI engineer, 4 weeks, monthly plan160Published 160-hour plan$6,100
One AI engineer, 6 weeks, hourly240240 × $45$10,800
6 weeks plus a senior AI engineer 1 day a week240 + 48240 × $45 + 48 × $60$13,680
One AI engineer, 4 weeks, with US Eastern overlap160160 × ($45 + 15%) = 160 × $51.75$8,280
UK: one AI engineer, 20 days16020 × £288£5,760
UK: one AI engineer, 30 days24030 × £288£8,640

For US teams, 09:00 to 12:00 US Eastern overlap runs through an evening shift at 15% extra. UK teams get 09:00 to 13:00 UK time every working day at no extra cost.

For comparison, IT Jobs Watch shows a £588 median contract day rate for UK AI engineer roles in the six months to 17 November 2026 (IT Jobs Watch), or £11,760 for 20 days. The US median annual wage for software developers was $135,980 in May 2025 (BLS): about $10,460 for four weeks ($135,980 ÷ 52 × 4), before benefits and hiring time.

Model fees are small by comparison: running a 300-case test set 20 times is 6,000 calls, or 6,000 × $0.011 = $66 at the mid-priced example above.

What moves the cost up:

  • Data preparation: cleaning, redacting or labeling records before any test can run.
  • Test set: a domain expert's time to write the correct answers.
  • Integrations: each live system adds access, error handling and testing.
  • Your team's time: an owner and reviewers who read outputs every week.

For what a full build costs once the PoC passes, see our guides to AI agent development cost and AI chatbot cost.

Why most AI pilots never reach production

Because the pilot proved the model could do the task, not that the organization could run it: data, workflow fit, ownership and budget were left for later.

SourceDateFinding
MIT NANDA, The GenAI Divide (Fortune)2025 reportAbout 5% of AI pilot programs achieve rapid revenue acceleration
MIT NANDA, same report (Virtualization Review)July 2025 report60% of organizations evaluated enterprise-grade systems, 20% reached pilot and 5% reached production
S&P Global Market Intelligence (Fortune)2025 survey of over 1,000 respondents42% of companies scrapped most of their AI initiatives, up from 17% in 2024
IDC research (CIO)March 2025For every 33 AI PoCs launched, only four reached production
RANDAugust 2024By some estimates, more than 80% of AI projects fail, twice the rate of IT projects without AI

The studies define failure differently, and RAND's excluded projects that only prompted pretrained LLMs. MIT's figures come from 150 interviews with leaders, a survey of 350 employees and 300 public AI deployments (Fortune). Read all of them as a direction, not a precise rate.

The reasons repeat across them:

  • Wrong problem: RAND's interviews with 65 data scientists and engineers found misunderstanding the problem to be the most common root cause; missing data and weak infrastructure are also on its list of five.
  • Workflow fit: MIT NANDA blames a "learning gap"; its lead author says generic tools "don't learn from or adapt to workflows" (Fortune).
  • Readiness: IDC said the low conversion "indicates the low level of organizational readiness in terms of data, processes and IT infrastructure" (CIO).
  • Cost and risk: S&P respondents named cost, data privacy and security risks as the top obstacles (Construction Dive).

Google engineers described the trap in 2015, warning that "time pressures may encourage a prototyping system to be used as a production solution", and that "results found at small scale rarely reflect the reality at full scale" (NeurIPS). Their first figure is captioned: "Only a small fraction of real-world ML systems is composed of the ML code."

From AI proof of concept to production: a checklist

Before calling a proof of concept a success, check that each line has an answer. A blank line usually means a stalled pilot, however good the scores.

  • Data access: a lasting, approved route to the data, not a one-off export.
  • Evals: the test set lives in your repository and runs on every change to a prompt, model or tool, as OpenAI's continuous evaluation advice says: "run evals on every change". Keep it in your own systems: OpenAI is retiring its hosted Evals platform, with shutdown scheduled for 30 November 2026 (OpenAI).
  • Monitoring: weekly samples of outputs, plus cost per task, latency, errors and misuse.
  • Model changes: Anthropic gives at least 60 days' notice before retiring publicly released models, and advises testing replacements "well before the retirement date" (Anthropic).
  • Security review: test against the OWASP list above. GOV.UK Chat also ran a data protection impact assessment and red teaming with government security experts (AI Playbook).
  • Human approval: for agents that act, the UK AI Playbook says "there must be a human present to review the action before carrying it out" (AI Playbook).
  • Owner: one named business owner who controls the workflow and answers for the results.
  • Budget: a monthly run budget for model fees, hosting and people time, approved before the pilot ends.
  • Exit rules: written criteria for scaling, changing scope or switching it off.

For a framework, the NIST AI Risk Management Framework (January 2023, voluntary) groups this work into four functions: govern, map, measure and manage (NIST AI RMF core). NIST added a Generative AI Profile on 26 July 2024.

How to choose the use case for your first AI proof of concept

Pick a task that happens often, has a checkable output and already has data behind it. Start internal, where a wrong answer is caught before a customer sees it. RAND's advice is to "focus on the problem, not the technology" (RAND).

  • Volume: the task happens hundreds of times a month, so small gains add up.
  • Checkable: a person can say whether an output is right, which makes a test set possible.
  • Data: the documents or records exist and you are allowed to use them.
  • Simplest fit: Anthropic recommends "finding the simplest solution possible", which "might mean not building agentic systems at all". A fixed workflow with one AI step is often enough for a first PoC.

If you have several candidates and no clear winner, our free AI assessment looks at how your business runs and returns a roadmap of high-return AI opportunities, which is a practical way to choose your first PoC.

Where Webify.AI fits

We are Webify.AI, an AI development company headquartered in Ahmedabad, India, so we have an interest in this topic. We work mostly as an extension of your team rather than a consultancy: our engineers join your team, or we build a defined project.

  • Rates: AI engineers cost $45 an hour (£36) and senior AI engineers $60 (£48), or $6,100 and $8,150 for 160-hour monthly plans. US Eastern overlap (09:00 to 12:00, through an evening shift) costs 15% extra; UK overlap (09:00 to 13:00 UK time every working day) is at no extra cost.
  • Contract: we sign an NDA before the first call, you own all the code with IP assigned to you, developers are named in the contract and never swapped without your agreement, and there is no subcontracting without your consent.
  • Trial: a two-week trial with the developer you choose; no charge if you stop within the first five working days.
  • Limits: we hold no ISO 27001, SOC 2 or Cyber Essentials certification, and our engineers work remotely from India. Our security page explains what we do instead.

Frequently asked questions

Can an AI proof of concept be done in one week?

A demo can. A PoC with a test set, pass mark, cost and latency figures and a failure review rarely fits in a week, because collecting and labeling real cases takes time. It can work when a labeled test set exists and access is ready on day one.

How many test cases does an AI proof of concept need?

Enough to cover the real mix of inputs, including the rare and awkward ones. The worked evaluations in Anthropic's guide range from 50 groups of paraphrased questions to 1,000 labeled tweets, most at 100 to 500 cases, and its advice is to "prioritize volume over quality" where grading can be automated.

Keep part of the set back for the final run, so the score you report comes from cases nobody tuned the prompts against.

Can we use real customer data in a proof of concept?

Usually yes, because clean synthetic data hides the problems you need to find. Use the minimum needed, redact what you can, keep it in your own accounts and get data protection sign-off first. For GOV.UK Chat, the team ran a data protection impact assessment and removed personal data from pages the model could see (AI Playbook). Our guide on sharing data with an offshore AI team covers the contract and access side.

Can the proof of concept code go into production?

Sometimes, but plan for rework. The test set, prompts and findings carry forward; the code usually needs error handling, access control, logging and monitoring first. Keep everything in your own repositories, with the IP assigned to you in the contract.

What if the proof of concept misses its pass mark?

Then it did its job at the lowest cost. Choose a narrower task, a different approach (such as a human review step), or stop. Do not lower the pass mark after seeing the results.

Who should own an AI proof of concept?

One business owner who controls the workflow, signs off the pass mark and holds the production budget, working with a technical lead. RAND found that misunderstandings about "the intent and purpose of the project are the most common reasons for AI project failure" (RAND).

Not sure what to build first?

Our free AI assessment looks at how your business runs and returns a roadmap of the AI opportunities with the highest return. When you are ready to build, our AI engineers start from $45 an hour (15% extra for US Eastern hours).

Get a free AI assessment

Hire AI developers