Phased vertical agents: 3.5x more ROI than full deployment

There's a pattern that repeats every time an SME sits down to talk about AI with me. They arrive with a big idea: "we want an agent that manages the entire sales process." They leave the meeting with a budget, deadlines, and a six-month project. Nine months later, the system is half in production, the team uses it only occasionally, and nobody is able to defend the ROI to the board.

It's exactly the pattern Gartner documents: more than 40% of agentic projects will be cancelled before 2027 due to rising costs and unclear business value. And WRITER's figure is even harsher: only 23% of companies that deploy AI agents capture significant ROI.

It's not a technology problem. The models work. The problem is method.

The figure that changes the conversation

Research by Deloitte —corroborated by independent studies from Zendesk in 2025— concludes something worth reading slowly: organizations that implement AI in phases report 3.5 times more ROI than those that bet on full deployments from the start.

It's not that the phased approach is prettier. It's that the "all at once" approach is statistically a worse investment.

The reason is simple: in an agentic project, each phase teaches you what to build in the next one. Without that information, what you build isn't a solution, it's a six-month hypothesis with a budget.

How we do it at E2D: the August Collection case

I'm going to tell you about a real case we're building right now, because it illustrates the method better than any theory.

August Collection is an online gallery with collector's pieces ranging from €1,500 to €6,000. The client came with the usual question: "we want to automate the entire sales process with AI".

My answer was: no. Or rather: not yet.

We build in three phases, each with its own KPI and a decision point before continuing.

Phase 1 — In production now: per-artwork information assistant. Each listing has its own assistant that knows the context of the piece, author, technique, historical period, and answers specific questions at the moment the buyer is looking at that artwork. KPI: time on listing, conversion ratio to contact form, qualitative quality of the leads reaching the gallerist. Cost and time: a fraction of what the full project would have cost.

Before moving to the next phase, we need that phase to deliver value on its own. If it doesn't, we don't build on top of it.

Phase 2 — In development: automatic meeting scheduling. When the assistant detects high-intent signals —questions about dimensions, authentication, payment methods— it offers to schedule an in-person visit or video call with the gallerist, integrating with the calendar and notifying them with the prior context of the conversation. This only makes sense if phase 1 proves there are qualified leads that deserve that next step.

Phase 3 — On the roadmap: sale closing and reservation management. The agent generates a proforma invoice, manages a 48-hour reservation of the piece, follows up if the buyer stops responding, and escalates to a human according to clear rules (pieces above a certain threshold, doubts about authenticity, recurring buyers). Building this today, before validating the two previous phases, would be designing a solution to a problem we don't yet fully understand.

Each phase is a vertical agent that is functional on its own. And each phase validates whether the next one is worth the investment.

Why this method works and the other doesn't

1. It reduces risk to manageable fractions. If phase 1 doesn't deliver value, you stop and you've invested a small percentage of the total budget. With an "all at once" approach, you discover the problem when you've already spent 80%.

2. It turns ROI into something demonstrable, not projected. In each phase you have real data on the system's behavior with your real users. Phase 2 sells itself if phase 1 worked. And it's cancelled without drama if it didn't work.

3. It lets the system learn from your business before making critical decisions. An agent that closes sales needs to have first observed how the buyer behaves, what doubts they have, what signals precede a real purchase. That isn't known in a meeting room. It's known in production.

4. It creates real ownership on the client's side. When the process owner sees phase 1 working before approving phase 2, it stops being an IT project that crosses their desk. It becomes their project. This is the factor that separates the 12% who reach production from the 88% who stay in pilot.

5. The integration —which is where 46% of projects get stuck— is built gradually. Connecting an agent with a single system (website + product listing) is manageable. Connecting it with CRM + calendar + invoicing + WhatsApp + ERP all at once is where projects die.

What this means if you're evaluating whether to invest

If your potential partner brings you a six-month proposal to "build the complete vertical agent for your business", ask them two things:

• Which phase do you deliver in the first 6 weeks and what KPI will we use to decide whether to continue?
• What happens if that phase doesn't work as we expect? Do I have the option to stop without losing what I've invested?

If they can't answer with specifics, what you're buying isn't a vertical agent: it's a traditional software project with an AI label.

The 2026–2027 window is real —AI Act in August, Verifactu and B2B electronic invoicing before July 2027—, and that puts pressure on moving fast. But "fast" doesn't mean "all at once". It means starting phase 1 now, with a contained case, a clear KPI, and the real option to iterate.

The 23% that captures ROI isn't the fastest to deploy. It's the one that validates each step before taking the next.

How to assess if you're ready for phase 1

Before investing a single euro, check:

1. Do I have a repetitive process, with available data and a prior numerical KPI? Without a baseline, there's no way to prove that phase 1 worked.
2. Is there a process owner —not a committee— who commits to using the system and reporting results?
3. Do I accept that phase 1 will be deliberately contained, not the complete solution I imagined?

If you answer yes to all three, there's a case. If you fail on the third, the problem isn't the technology: it's the expectation.

At E2D we build vertical agents in phases, with KPIs defined before each deployment and the real option to stop if the evidence doesn't support it. If you want an honest conversation about whether your case is one to start now or wait, write to me at hello@evolve2digital.com.