The 90-Day Path From Data to a Managed AI Workforce

Ninety days is the number a vendor gives you when they want the deal to feel safe. It sounds disciplined. It fits a quarter. It also has almost nothing to do with the technology. You can stand up an agent that drafts emails in an afternoon. What takes 90 days is everything around it: getting the data clean enough to trust, deciding what the agent is allowed to touch without a human, and proving the thing moves a number you report to the board.

So this is not a countdown to a launch. It is a path to a managed AI workforce — agents that run in production, get supervised, get corrected, and stay tied to an outcome after the people who built them have gone home. The launch is the easy middle. The hard parts are the start and the part after. Here is how the 90 days really break down for the person who has to sign for it.

Why the clock is set by data, not by the model

The frontier models are commodities now. Anthropic, OpenAI, and the model inside Agentforce are all good enough to draft an email, qualify a lead, or summarize a case. None of that is your bottleneck. Your bottleneck is that the agent has to act on your reality — your accounts, your products, your pricing, your definition of a qualified lead — and that reality is scattered across a CRM, a billing system, a support tool, and three spreadsheets, with duplicate records and fields nobody has updated since 2023.

An agent fed bad data does not fail loudly. It fails confidently. It tells a customer the wrong renewal date, routes a deal to the wrong rep, or promises a discount that does not exist — and it does it in fluent, plausible language a reviewer is inclined to believe. That is why we say data before agents, every time. The model is rented. The data is yours, and it is the only part that decides whether the agent is an asset or a liability you cannot see until a customer screenshots it.

“

The model is the cheapest part of the system. The data and the accountability are the expensive parts, and they are the parts no vendor wants on the invoice.

Days 1–30: Agent Ready — make the data tell the truth

The first month buys nothing visible. No demo, no dashboard, no win you can show the board. It is the month where you pick the single workflow worth automating and make the data under it trustworthy. Concretely, on Salesforce Data Cloud that means resolving records into one profile per customer with an identity-resolution ruleset you actually agree with, mapping the source fields the agent will read, and writing down the three or four definitions it will rely on — what counts as a qualified lead, an active account, a churn risk — so the agent and the business mean the same thing by them.

The question for this phase is not "is it built yet." It is:

• What is the one workflow where a fast, correct response is worth real money, and how much money?
• What records and fields does the agent need to read to do that job, and are they clean and current right now?
• What is the agent allowed to do on its own, and which actions require a human to approve before they commit?
• How will we know, in dollars, whether it worked — and what is the baseline today, measured, not guessed?

If you cannot answer the last one before launch, you will not be able to answer it after, and you will have bought a science project. Pin the baseline now. "We respond to inbound leads in a median of nine hours and book a meeting on 4% of them" is a sentence you can hold an agent accountable to. "We want to be more efficient" is not. The numbers there are an example — yours come from your own pipeline, and pulling them is part of the first 30 days, not an afterthought.

Days 31–60: Agent Launch — narrow, supervised, in production

The second month is the one that feels like progress, and it is the one where teams most often overreach. The discipline is to launch the agent narrow and watched. One workflow. A bounded set of actions defined explicitly, not left open. A human in the loop on anything irreversible — anything that touches money, a contract, or a customer's trust. Everything the agent does is logged with the data it saw and the action it took, so a bad outcome is debuggable instead of mysterious.

Take speed-to-lead, the kind of work we ran for a residential solar client through Green Subsidy. The agent's job was narrow and valuable: respond to a new inbound lead in the moment instead of hours later, qualify it against agreed criteria, and book the next step on a human's calendar. One workflow, a bounded action set, and a metric — response time and booked appointments — that reads without translation. You do not need the agent to do ten things in month two. You need it to do one thing reliably and prove it on real traffic, not in a sandbox where every input is friendly.

"In production" is the operative phrase. A pilot that never touches a real customer proves only that the demo works on the demo's data. The 60-day mark should have an agent handling live volume under supervision, with every action logged and a one-click path to roll back the scope if it misbehaves. That is what makes the number at day 90 a measurement instead of a hope.

Days 61–90: Agent Scale and Agent Care — the part the demos skip

An AI agent is not software you install and forget. It is closer to a new hire than a new app. The business changes around it — a new product, a pricing update, a policy shift — and the agent keeps confidently doing the old thing until someone notices. Left alone, an agent that was right in March is quietly wrong by June, and the failure mode is silence, not an error log.

So the last 30 days are not a victory lap. They are where the managed in managed workforce gets earned. You widen scope only where the agent has a track record. You watch the cases it handles badly — the edge inputs, the angry customer, the deal that does not fit the criteria — and you route those to people on purpose. You update its instructions and its data when the business moves. This is ongoing work, what we call Agent Care, and it is the single biggest reason internal AI projects stall after a strong pilot. The team that built it moves to the next thing, nobody owns the running system, and the agent rots in place while still answering customers.

“

Most failed agent projects did not fail at launch. They failed around week twelve, when the people who built the agent moved on and nobody was left to run it.

The honest version of "90 days"

If a partner promises you a finished, autonomous AI workforce in 90 days, read it as a sales claim, not an engineering one. What 90 disciplined days realistically gets you is this: clean data under one valuable workflow, one agent running that workflow in production under human supervision, and a measured result against a baseline you set on day one. That is a genuinely strong outcome. It is also the beginning of the operating relationship, not the end of the project — and a partner who tells you otherwise is the one to be careful with.

The companies that win with agents treat this less like a software purchase and more like hiring. You would not hire a salesperson, hand them no accounts, give them no manager, and check on them once a quarter. An agent needs the same things a good employee needs: clean information, clear boundaries, a supervisor who reviews the work, and a number it is responsible for.

What a CFO should actually underwrite

Strip away the model talk and the underwriting decision is plain. You are buying down a known cost or buying up a known revenue, against a baseline you can defend in a board meeting. Frame it as an illustration, not a promise: say you book meetings on 10% of inbound leads today, and an always-on response that reaches them in seconds lifts that even a couple of points — you can size that against your own lead volume and average deal value before you sign anything. If the math only works on a number nobody in the room can name, that is your signal to wait.

This is why we tie our fee to the return rather than to the build. It changes who carries the risk. A build-and-bill vendor gets paid whether the agent earns its keep or not; their incentive ends at go-live, which is exactly when yours begins. When the fee is tied to the outcome, the people who built the agent are the same people on the hook to keep it working in month four and month ten. Build it, run it, stay accountable for it — the accountability is the part that actually protects the buyer.

Run the 90 days as three questions, in order. Is the data true? Is the agent live and watched? Is the number moving? If you can answer yes to all three, you do not have a pilot. You have a managed AI workforce and a baseline to grow it from.