Paul Jialiang Wu agentic-portfolio 中文 Español 한국어 日本語✉️ Free list
← Back to portfolio

AI-NATIVE SERIES · PITCHING AGENTIC AI, PART 2 OF 3

Selling Agents This Season: Bounded, Accountable, Evidence-Led

By Paul Jialiang Wu · agentic-portfolio-lovat.vercel.app · 2026-10-21 · 6 min read

Cover: Selling agents this season, bounded, accountable, evidence-led. Three cards: who it is for (founders, consultants and platform teams pitching an agent this year), what you leave with (six practices and the four checks every agent pitch should pass), and the one finding (developers felt 20% faster with AI but were 19% slower, METR 2025). Series map along the bottom, part 2 highlighted.
Pitching Agentic AI, part 2 of 3.

1-minute takeaway

What this is. Part 2 of the series: this season's practice for pitching AI agents, and four checks every pitch should pass: specific, sized honestly, owned by a person, measured in advance.

Why it matters. In a 2025 randomised trial, experienced developers were 19% slower with AI tools and still believed they were 20% faster (METR). A broad productivity promise is the claim the evidence supports least.

What you can do after reading. Run your four lines through the four checks, then write one sentence for the person who can say no: what the agent can touch and who approves.

Pitching Agentic AI — the series spine
  1. Principles that outlast every pitch episode 1
  2. Selling agents this season this episode
  3. Four clients, four first workflows episode 3

How to read the labels: EVIDENCE is sourced, DESIGN is a recommended practice, OPEN is not yet tested, and LENS is my own conviction.

Enterprise AI pilots often stall, and how they're scoped is a large part of why. Six practices for 2025–26, built around one idea: pick one stuck workflow, and let a paired measure agreed in advance do the selling. The surprise: the best trials of AI coding tools disagree.

What this episode covers: how to pitch an AI agent right now

Part 1's principles barely move. This layer moves every year or so: how buyers buy, what the evidence currently says, and what they ask about governance. It is for anyone pitching an agent into an enterprise this year, whether a startup founder, a consultant or an internal platform team. I work at Accenture, which appears in one of the studies below; I say so where it comes up. Everything here is my own synthesis from public sources.

Top: the four moves in order, Listen, Locate, Limit, Prove, joined by arrows. Below: three bars for the series layers, principles, this season and your client, shortest last. Part 2, this season, is highlighted.
Every part uses the same four moves. The shorter the bar, the sooner that layer needs re-checking.

Make a bet first

In a 2025 randomised trial, experienced open-source developers worked on real tasks in their own repositories, with and without AI tools. Before starting they predicted AI would make them about 24% faster. With AI, their tasks took:

Practice 1: qualify before you pitch

EVIDENCE Enterprise sales formalised this long ago. MEDDIC, a checklist practitioners trace to PTC in the 1990s, asks for metrics, an economic buyer, decision criteria, a decision process, the pain, and a champion. Its origin story is practitioner lore and unverified in detail, but it survives for a plain reason: a deal with no owner and no agreed measure rarely closes, however good the demo.

DESIGN For an agent trial I add four questions of my own. Is there a named owner for the result? Can the team give access and context? Is there a reviewer with time to judge the output? Is the work repeatable enough to scope, with a way to check it? Missing one is a reason to pause and find it, not to push on.

Practice 2: pick one workflow under pressure

DESIGN Stuck engineering work tends to look like one of four things. I call them the four Rs:

| Stuck work looks like | Typical symptom | A bounded first task | |---|---|---| | Waiting | reviews and approvals queue up | first-pass review against the team's own checklist | | Redoing | rework, flaky tests, re-run pipelines | triage failing CI runs for one service | | Relearning | context lost between people, old code nobody knows | explain and document one legacy module, read-only | | Re-explaining | handoffs, stale docs, repeated questions | draft a runbook from one incident's history |

Pick one row, for one team. Not as narrow as autocomplete, and not as broad as company-wide automation, which is just slide 7 with a budget.

Practice 3: speak to each person who can say yes, or no

DESIGN The same stuck workflow means something different to each of four people:

Most pitches address the first three. In the deals I've watched stall, it was usually the fourth person, who never got a sentence.

Practice 4: claim only what the evidence supports

EVIDENCE Back to the bet: the answer is C. In METR's 2025 randomised trial (16 experienced developers, 246 tasks in their own repositories), allowing AI tools made tasks take 19% longer, and the developers still believed afterwards that AI had made them about 20% faster (Becker et al., METR, 2025). METR later said a re-run gave only a weak signal, because many developers would no longer work without AI, and that it now thinks developers are likely sped up more than before (METR, Uplift update, Feb 2026).

EVIDENCE The other side: field experiments at three companies, including Accenture (my employer), with 4,867 developers found about 26% more completed tasks (pull requests, not "productivity") with an AI code-completion tool, with less experienced developers gaining more (Cui et al., The Effects of Generative AI on High-Skilled Work, 2024–25). Google's DORA 2025 survey of about 5,000 professionals found AI use tied to higher throughput but still to lower delivery stability: AI amplifies what's already there (DORA 2025 announcement).

DESIGN These don't contradict each other once you read them by setting and by era: code completion for a broad population in 2022–23, experienced developers in mature codebases with early-2025 tools, and whole delivery systems. That is why "developer productivity" is the wrong promise, and a workflow-specific claim with an agreed measure is the honest one.

Practice 5: pair a speed signal with a quality signal

EVIDENCE Whether a pilot reaches production depends heavily on how it is built. MIT NANDA's 2025 study, based on interviews, surveys and public implementations, reported that "external partnerships with learning-capable, customized tools reached deployment ~67% of the time, compared to ~33% for internally built tools," and itself notes the figures are self-reported (MIT NANDA, The GenAI Divide, 2025; a widely circulated copy of the report). Its often-quoted "95%" counts organisations getting zero return, not failed pilots.

EVIDENCE Vendors are moving pricing towards outcomes: Intercom charges $0.99 per resolution for its Fin agent (Intercom pricing), and Salesforce prices Agentforce per conversation or per action (Agentforce pricing). Each vendor defines "outcome" itself.

DESIGN This is the Prove move with teeth. Agree a speed signal and a quality signal together, with the bar and who reads them: review turnaround with escaped defects, documentation time with tests accepted unchanged. Speed alone is gameable. Reviews get faster by getting shallower, and everyone celebrates until the incident review.

Practice 6: let the trial's shape answer governance

EVIDENCE Buyers now arrive with a governance checklist. The EU AI Act has applied obligations for general-purpose AI models since August 2025 (EU AI Act implementation timeline), and NIST's AI Risk Management Framework is the common US reference (NIST AI RMF).

DESIGN A limited trial answers most governance questions by its shape. Say what the agent can touch, who approves each output, what is logged, and when a specialist will check configuration and controls. If the answer to "what can it touch?" takes a paragraph, the trial is too big.

Four checks before you pitch

  1. Specific: one workflow, one team.
  2. Sized honestly: "may be worth evaluating", not a promised transformation.
  3. Owned by a person: engineers keep responsibility for quality and approvals.
  4. Measured in advance: a speed and a quality signal, with a bar, agreed before starting.

How this could fail

Your move this week

Take the four lines you wrote in part 1. Run them through the four checks. Then write one sentence for the person who can say no: what the agent can touch and who approves. If you can't write it, you aren't ready to propose a trial.

Next

Part 3 tailors all of this for four industries where mistakes are most expensive: financial services, life sciences, industrial manufacturing, and technology and telecom. Each gets the pain, the owner, the rules, and a first workflow that fits.

Sources


Related

Written by Paul Jialiang Wu · agentic-portfolio · 2026-10-21. Views are my own, not my employer's. This page is generated from a gated source: length, sources, claim labels and the honest status are checked before it can ship.