An AI implementation roadmap: from first use case to production in nine weeks

An AI implementation roadmap is a phased plan that takes one business problem from idea to a working system your team uses every day. For a focused first use case with accessible data, about nine weeks from the first workshop to production is a realistic target: two weeks of discovery, a week to prioritize, two weeks to prove the approach on your data, and four weeks to build, connect and roll it out.

Most companies do not struggle to find ideas for AI. They struggle to turn one of them into software that runs reliably inside real work. This guide sets out the phases we use at DeepVelocity, what each one should produce, the decision at the end of each, and the mistakes that most often stretch a few weeks into a year.

Select a phase in the timeline below to see what it involves, and add common complications to see which phases they stretch.

Before you start: one use case, one owner, one number

The most reliable way to implement AI in a business is to start smaller than feels ambitious. Three things should be true before the first workshop:

• One use case: a specific workflow or step, such as drafting quotes, sorting inbound requests or forecasting demand for a product line. "Use AI across operations" is a direction, not a use case.
• One business owner: a person who runs the work today, cares about the outcome and can make decisions quickly. Without an owner, projects drift.
• One number: the measure that will show whether it worked, with today's value as the baseline. Hours per quote, response time, error rate or revenue per rep.

If you do not yet know which use case to start with, that is exactly what the first two phases are for.

Phase 1: Discovery (weeks 1 to 2)

Goal: understand how the work is done today and where AI could change it.

Discovery is workshops and observation, not a questionnaire. Sit with the people who do the work, map each step, and note where time goes, where errors happen and where decisions wait on information. At the same time, look at the data: where it lives, who owns it, how complete it is, and whether you are allowed to use it for this purpose.

• Who is involved: the business owner, two or three people who do the work, someone from IT or data who knows the systems, and the AI team.
• Deliverables: a map of the current workflow, a long list of opportunities, and a first read on data availability and quality.
• Decision at the end: is there at least one opportunity worth sizing, with data that could plausibly support it?

Common mistake: talking only to managers. The people doing the work know where the time really goes, and which "simple" steps are full of exceptions.

Phase 2: Value at stake and prioritization (week 3)

Goal: choose the first use case on evidence, not enthusiasm.

Size each opportunity in your own terms: hours saved, revenue gained, errors avoided or faster turnaround. Then score its feasibility. Is the data there? How many systems does it touch? How costly is a mistake? How much will the workflow have to change? Plot value against feasibility. The best first project sits high on both. Valuable but difficult projects belong later in the roadmap, once you have better data, the integrations in place and some experience.

How to choose the first AI use case

Good first use cases tend to share a few traits:

• The task is repetitive and frequent, so small gains add up.
• The inputs are mostly digital: documents, emails, records and forms.
• A person can check the output easily, at least at first.
• Success can be measured with a number you already track.
• It matters to someone senior enough to clear obstacles.

• Who is involved: the business owner, a senior sponsor, finance if savings are being sized, and the AI team.
• Deliverables: sized opportunities, a ranked shortlist, the chosen first use case with its success measure and baseline, and a rough plan and budget for what follows.
• Decision at the end: do we commit to a proof of concept on this use case?

Common mistake: picking the most impressive idea rather than the most feasible one. A first project that ships builds the case for the second. One that stalls can set the whole effort back by a year.

Phase 3: Proof of concept on your own data (weeks 4 to 5)

Goal: show the approach works on your real data, well enough to justify the full build.

A proof of concept is not a demo. A demo shows what AI can do in general, usually on tidy examples. A proof of concept runs on your documents, your records and your awkward cases, and is judged against a test set agreed before any work starts: a set of real examples with known right answers, chosen by the people who do the work. It answers practical questions. How accurate is it? Where does it fail? What will it cost to run for each item processed? How will people review its output?

• Who is involved: the AI team, with the people who do the work choosing test cases and judging the results.
• Deliverables: a working prototype on your data, results against the agreed test set, a list of failure cases and how to handle them, estimated running costs, and a design for the production system.
• Decision at the end: go, adjust or stop.

Stopping at this point is a good outcome, not a failure. It costs two weeks instead of a full build, and you learn something about your data either way.

Common mistake: starting without an agreed definition of "good enough", so the result gets debated instead of decided. Our guide to why AI pilots fail covers this and the other traps that keep projects stuck in testing.

Phase 4: Production build (weeks 6 to 9)

Goal: turn the proof of concept into software your team relies on every day.

This is where most of the engineering happens, and where many AI projects stall. A production system has to connect to the tools people already use, handle errors and unusual inputs gracefully, keep a record of what the AI did and why, protect sensitive data, and be monitored once it is live. The human steps need designing too: who reviews what, how they correct the AI, and how those corrections lead to improvements.

Roll out in stages. Start with a small group using the new system alongside the old process, compare results, fix what they find, then widen it to the rest of the team.

• Who is involved: the AI team, IT and security, the business owner and a pilot group of users.
• Deliverables: a production system connected to your tools, monitoring and alerts, documentation and training, and a staged rollout plan.
• Decision at the end: does the pilot group's result meet the success measure, and is the team ready to switch over?

Common mistake: leaving integration work and security review until the last weeks. They usually take the longest to arrange, so start them during the proof of concept.

Phase 5: Maintenance and change management (ongoing)

Goal: keep the system working, keep it improving, and make sure people use it.

An AI system is not finished at launch. Volumes grow, new kinds of input appear, your processes change, and AI model providers retire and replace the models underneath. Someone has to watch quality, fix issues and update the system as the business moves. Whether that is your team or the firm that built it is a staffing and pricing decision, covered in our guide to AI consulting pricing models.

Adoption needs as much attention as the software. The habits that make it stick:

• Involve users from discovery onward, so the system reflects how they actually work.
• Train on real cases from their own work, not on slides.
• Make it easy to flag a wrong answer, and show people that their corrections lead to fixes.
• Retire the old process on a set date, so two ways of working do not run side by side forever.
• Report the success measure regularly, against the baseline, to the people doing the work as well as to leadership.

How long does AI implementation take?

For a focused first use case with accessible data and modern systems, about nine weeks is realistic. The timeline grows when:

• The data needs work: records spread across spreadsheets, inboxes and old systems have to be gathered, cleaned and sometimes labeled before anything can be tested.
• Older systems are involved: software without modern interfaces takes longer to connect to.
• Reviews are heavy: security, legal and compliance approvals, especially for sensitive or regulated data.
• The scope grows: each extra workflow step or department adds design, testing and training.
• Decisions are slow: if the business owner cannot make calls quickly, every phase waits longer at its decision point.

When a plan runs long, cut scope before adding time. A smaller system in production is worth more than a larger one still in testing, and it gives you real results to build on.

How to measure results after launch

Track the number you chose in phase 2 against its baseline, alongside a few health measures: how often people accept the AI's output without changes, how often they correct it, how long items wait for review, and the running cost for each item processed. Review them monthly with the business owner. A falling acceptance rate is often the first sign that something has changed, such as a new type of input, a change in the process or an update to the underlying model.

After the first use case: build a portfolio, not a program

Once the first system is in production, go back to the ranked list from phase 2. The second project is usually faster: the data is better understood, some integrations already exist, and the team knows what to expect. Over time the roadmap becomes a portfolio of focused systems, each with an owner and a measured result, rather than one large program that has to succeed all at once.

Not every opportunity on the list needs a custom build. Some are better served by an off-the-shelf product, and some by an agent that handles a routine task end to end. Our guides to custom AI versus off-the-shelf tools and AI agents for business explain how to tell which is which.

How DeepVelocity runs AI implementation

This roadmap is how we work with clients: discovery, sizing the value at stake, a proof of concept on your own data, a production build, then maintenance and change management. Our founders, both formerly of McKinsey & Company, have built and deployed AI-powered products and supported technology transformations at Fortune 500 and growth-stage companies. That combination is what lets us move from discovery to production in weeks and months, not years. See how we approach custom AI solutions and machine learning models trained on your data, or tell us about your first use case.