Your mid-market company is leaking EBITDA every quarter. Agentic AI recovers it — and you deploy the system in weeks, not quarters.

The evidence is stark. MIT's 2025 study The GenAI Divide, conducted by its Project NANDA initiative across 300 public deployments, 150 leader interviews, and 350 employee surveys, found that 95% of enterprise generative-AI pilots delivered no measurable P&L impact, while only 5% captured significant value. Its central conclusion is the part operators should internalize: the gap is not explained by model quality but by a “learning gap” — the failure to integrate AI into real workflows. Gartner has separately projected that more than 40% of agentic-AI projects will be scrapped by 2027.

The mechanism is consistent. When AI is run as a technology initiative, success is defined as “go-live” — a model shipped, a dashboard built — and the program produces activity that never reaches margin. The minority that succeed invert the sequence: they begin with a financial baseline, ask where margin is actually leaking, attach each agent to a specific metric, and refuse to scale any phase that has not cleared a measured threshold.

That discipline has a name — gating. In the SEAS framework's modeled deployment, a Phase-1 gate at roughly week eight requires defined thresholds — accuracy above 90%, cycle-time reduction above 25%, cost reduction above 15%, exception rate below 8%, and user satisfaction above 3.5/5 — all of which must be met before scope widens. A pilot that cannot clear the gate is paused, not propagated. This is simply the operational expression of MIT's finding: integration and accountability, not algorithms, decide the outcome.

Public benchmarks bound the opportunity. APQC's process data shows top-quartile accounts-payable teams operating at roughly $2–3 per invoice while bottom-quartile peers exceed $10 — a four-to-fivefold gap that recurs across back-office processes and represents recoverable cost, not theoretical savings. BCG's 2025 research frames the upside from the other direction: its “future-built” AI leaders carry 1.6 times the EBIT margin of laggards. The recoverable amount, in practice, is the distance between a company's current process performance and that demonstrated frontier.

In the SEAS reference model — a $150M-revenue, $18M-EBITDA company at a 12% margin — that distance totals approximately $5.3M per year, or about 3.6 margin points. It is not evenly spread. It concentrates in supply-chain and vendor spend (~$2.5M recoverable of ~$6M identified), administrative and SG&A processes (~$1.8M of ~$4M), and energy and operational overhead (~$1M of ~$2M).

The rigor is in underwriting only the portion you can tie to a baseline and sustain — the bankable subset — rather than the theoretical maximum. A board underwrites a defended number, not an aspiration. Held, $5.3M of recovered EBITDA at a typical exit multiple compounds: in the modeled case it adds meaningful enterprise value and roughly a full additional turn of MOIC at exit, before any multiple re-rating is counted.

The first pocket is SG&A and administrative process inefficiency — manual, exception-heavy back-office work such as accounts payable, reconciliations, and reporting that quietly inflates cost per transaction. APQC's data shows bottom-quartile AP functions cost four to five times the top quartile for the same task. The second is pricing and contract drift — margin lost to unmanaged discounting, unenforced terms, and renewals that are never repriced. The third is supply-chain and vendor concentration — overspend hidden in single-source dependencies and maverick buying, plus working capital trapped in extended days-sales-outstanding.

Each pocket is invisible on a standard P&L because it is spread across hundreds or thousands of small transactions, none material on its own. Traditional analysis samples; it cannot examine every invoice, contract line, and purchase order. This is precisely the structural advantage of agentic AI: it can interrogate the full transaction population rather than a sample, which is why it surfaces leakage that periodic audits and consulting reviews miss. It also explains why MIT found value accrues to workflow-level integration — the leakage lives in the workflow, not in the headline numbers.

The disciplined first move is therefore not deployment but diagnosis: a structured assessment that scores all three pockets against external benchmarks and produces a baseline before any technology is purchased.

The distinction is technical, not semantic. RPA executes a fixed workflow — it repeats the same interface actions a human demonstrated — and by definition is not artificial intelligence; it breaks the moment a transaction departs from the rule. Its strength is the predictable portion of a process; its ceiling is the exception. And the exceptions are exactly where cost and leakage concentrate: APQC's benchmarks show even strong AP functions running roughly 9% exception rates and weak ones around 22%, and exceptions are the expensive, manual, error-prone work.

RPA cannot touch that remainder — it automates the cheap majority and leaves the costly minority. Agentic AI is built for it: it reads unstructured documents, makes context-dependent decisions, escalates only genuine exceptions, and improves from feedback. This is the “intelligent automation” layer that pairs reasoning with execution rather than scripting alone.

The implication for deployment is direct. A program built only on RPA plateaus at the easy tasks — consistent with the long-reported difficulty of scaling RPA beyond initial pilots — while an agent-based approach reaches the judgment-heavy work that actually moves margin, under human oversight. The right model is not “AI instead of people,” but agents owning high-volume judgment work while people supervise exceptions and policy.

The principle is well established. Robert Cooper's stage-gate methodology, proven in new-product development for decades, holds that investment should advance through discrete phases, each ending in a go/no-go decision against predefined criteria. Applied to AI, it directly counters the failure mode MIT identified — it forces every phase to prove a financial result before the next is funded.

A defensible sequence has four moves. Diagnose: quantify where EBITDA is leaking and set a baseline. Pilot: deploy against one or two high-value processes, contained. Gate: validate the pilot against hard metrics before any expansion — in the SEAS model, a Phase-1 gate around week eight requires accuracy above 90%, cycle-time reduction above 25%, cost reduction above 15%, exception rate below 8%, and satisfaction above 3.5/5, all met. Scale: extend to further processes, each behind its own gate.

This is what protects the CFO. A failing pilot is paused, not propagated, so capital is never committed to an unproven approach; and each gate produces a documented, board-defensible track record. It is also why staged deployments succeed where big-bang programs stall — a conclusion BCG and McKinsey reach independently in finding that value comes from disciplined, end-to-end transformation rather than scattered pilots.

The problem with most reporting is measurement, not effort. McKinsey's research captures the consequence: 88% of companies use AI in at least one function, but only 39% can point to any EBIT impact, and usually under five percent. The activity is real; the attributable number is missing.

The method is to baseline each target process before deployment — AP cost per invoice, days-sales-outstanding, contract leakage — measure the delta afterward, and count only the sustained portion. “AP cost per invoice fell from X to Y, releasing $Z, validated over four weeks” is defensible; a productivity anecdote is not. Aggregated across the pockets, this is how a modeled ~$5.3M recovery becomes ~3.6 margin points.

Margin expansion is only half the case. BCG's 2025 study finds AI leaders deliver 3.6 times the three-year shareholder return of laggards, and McKinsey finds digital and AI leaders outperform on total shareholder return by two-to-six times across sectors — evidence that the market re-rates operators who demonstrably run on AI. For a private-equity-held company, that combination — recovered EBITDA plus multiple re-rating — is what converts into enterprise value and improved MOIC at exit. The board does not buy technology; it underwrites a number with evidence behind it.