Hiring an AI consultancy should require three deliverables before any contract: a documented diagnosis of your situation, a roadmap with phases and measurable deliverables, and an ROI projection based on your actual processes. Without them, you're buying a generic promise, not a solution.
TL;DR: Don't sign without a diagnosis, a phased roadmap and a projected ROI. A serious consultancy delivers those three documents before the contract, not after.
Most companies hire AI consulting the same way they hired agencies in the 2000s: a polished deck, a vague case study from another industry and a proposal promising "up to 40% cost reduction", with not a single line on which process will change or how it will be measured.
The problem has a new name, but the mistake is the same: buying a solution before mapping the problem.
Why demand a documented diagnosis before the proposal?
A diagnosis is not a kickoff meeting. It's a document that names the processes that are candidates for AI, quantifies current volume (calls per day, tickets per month, analyst hours) and pinpoints the bottleneck.
Without a diagnosis, the proposal turns generic: "we'll deploy a customer service agent", but nobody knows whether the bottleneck is in customer service or in triage, whether the volume justifies AI or whether a better CRM workflow would be enough.
At Agência Rollin, the diagnosis maps three layers:
- Process: where the friction is (manual handoffs, rework, waiting).
- Volume: how many operations per day, how many people involved, average time.
- Estimated impact: how much time and cost can be freed up if that process runs on its own.
That mapping becomes the yardstick for ROI. If the consultancy doesn't deliver a diagnosis, it has no way to project a return, and you have no way to hold it accountable later.
The concrete deliverable of this phase is a 3- to 5-page document: a list of mapped processes, volumes, viable automation points and an effort estimate. Without it, don't move forward.
How does an implementation roadmap with measurable deliverables work?
A roadmap is not a schedule. It's the sequence of phases, each with a testable deliverable and a success metric defined before it starts.
A serious AI implementation splits the work into short cycles, usually four phases:
- Diagnosis and validation (1 to 2 weeks): maps processes, defines KPIs, projects ROI. Deliverable: diagnosis document.
- Proof of concept (PoC) (2 to 4 weeks): tests the automation at reduced scale, with real data. Deliverable: working prototype and performance report.
- Guided implementation (4 to 8 weeks): puts the solution into production, trains the team, handles edge cases. Deliverable: solution running + SOP (standard operating procedure).
- Monitoring and optimization (ongoing or per sprint): tracks metrics, refines prompts, adjusts thresholds. Deliverable: monthly report on realized ROI.
Each phase ends with a decision gate: the manager reviews the deliverable, compares it with the promised metric and decides whether to continue or stop.
A roadmap without per-phase metrics is theater: the consultancy "delivers" AI, but nobody knows whether it worked until three months later, when the contract has already been paid.
Our experience in AI consulting shows that 70% of projected savings already show up in the PoC: if the prototype doesn't deliver a measurable result in four weeks, the problem is in the diagnosis, not the technology.
How do you calculate ROI before signing the contract?
AI ROI is not "estimated future savings". It's a simple calculation: projected monthly gain ÷ total investment × 12 months.
To run the numbers, you need three real figures from your process:
- Current cost: how much the company spends today running that process manually (salary + overhead + time).
- Expected reduction: what percentage of that cost AI can take over (rarely 100%; usually between 30% and 70%).
- Investment: how much it costs to implement and maintain the solution in the first 12 months (consulting + infrastructure + adjustments).
A real example of WhatsApp customer service:
| Metric | Before | After (projected) |
|---|---|---|
| Calls/month | 3,200 | 3,200 |
| Answered by a human | 3,200 (100%) | 960 (30%) |
| Agent hours/month | 320h | 96h |
| Monthly cost (R$ 25/h) | R$ 8,000 | R$ 2,400 |
| Monthly savings | — | R$ 5,600 |
| Investment (implementation) | — | R$ 18,000 |
| Payback | — | 3.2 months |
If the consultancy doesn't bring this table filled in with your company's data, it's selling a generic case.
Agência Rollin's AI cost calculator lets you simulate scenarios before the sales conversation: the manager enters call volume, hourly cost per agent and type of service, and gets a projection of savings and payback.
A tool like this should be a prerequisite: if the consultancy gives you no way to test the math yourself, be suspicious.
What questions should you ask in the sales meeting before closing?
The right questions separate promises from delivery. Use these five:
- "Do you deliver a documented diagnosis before the commercial proposal?" If the answer is "we do that in the first paid phase", walk away.
- "What is the PoC success metric, and what happens if it isn't met?" A serious consultancy refunds or readjusts; it doesn't push you into phase 2.
- "Who will run the solution after deployment, you or my team?" If it's your team, demand training and documentation; if it's theirs, demand an SLA.
- "What is the recurring monthly cost after implementation?" AI infrastructure (API, hosting, storage) costs money. It has to be in the ROI.
- "Do you have a documented case in my industry with similar volume?" It doesn't need to be identical, but it needs to be comparable. A consultancy that only shows a bank case to sell to a clinic is improvising.
Write down the answers. If the consultancy can't answer on the spot or promises to "customize later", it hasn't done the diagnosis, and it will do it with your money.
A sixth, technical question is decisive: "Does the solution use a proprietary API, or can it switch LLMs?" Lock-in to a specific model (GPT-4 only, Claude only) doubles the risk of future costs. A model-agnostic architecture lets you switch models as price and performance change.
What changes when the consultancy offers ongoing support?
Ongoing support turns AI from a project into an asset. The difference lies in the pricing model and the monthly deliverables.
Traditional project: you pay for implementation, receive the solution, and three months later it's outdated (prompts not refined, new cases unhandled, metrics dropping).
Ongoing support: the consultancy delivers a monthly optimization sprint, adjusting thresholds, retraining with new cases, adding intents and monitoring performance drift.
The recurring cost (usually 10% to 20% of the initial investment per month) has to fit within the projected ROI. If monthly savings are R$ 5,600 and support costs R$ 1,800/month, net ROI drops, but the solution stays alive.
In practice, AI agents for businesses that run without ongoing support lose 15% to 25% of their accuracy within six months: users change the way they ask, the product changes, edge cases appear.
Ongoing support is an investment, not a cost, if it comes with a monthly report on realized ROI. Demand a dashboard: calls processed, resolution rate, fallback to a human, average time, accumulated savings.
Without a published monthly metric, support becomes phantom maintenance: you pay, but you don't know whether the AI still delivers the promised ROI.
Key takeaways
- Demand a documented diagnosis before the commercial proposal; it's the only way to project real ROI.
- The roadmap needs short phases (2 to 4 weeks) and a success metric for each phase, not just at the end.
- Calculate ROI with your own process numbers: current cost, expected reduction and total investment.
- The PoC is a decision gate: if it doesn't deliver a measurable result in four weeks, stop and review the diagnosis.
- Ongoing support keeps the AI alive, but it's only worth it if it comes with a monthly report on realized ROI.
Want to project AI ROI for your scenario before any meeting? Use the AI cost calculator and simulate savings, payback and recurring cost with your own operation's numbers. It takes three minutes and gives you the math ready to compare proposals.
