AI coding agents produced three different productivity numbers this year. All three are real. Only one belongs in a value creation plan.

One study measured experienced developers 19% slower while they believed they'd been 20% faster. Another found a 19.3% net average gain from a broad survey of engineering leaders. A third claims 3 to 5x productivity, with teams 60% smaller, from companies that rebuilt their entire delivery model around it.

Here's where each one actually belongs.

The METR study behind the first number is a randomized controlled trial, not a survey: sixteen experienced developers, two hundred forty-six real tasks on their own codebases, AI allowed on half of them at random. The developers who used AI took longer to finish. The same developers, afterward, believed they'd been faster. That gap between what got measured and what got felt is most of this issue's argument.

Gartner evaluated twelve vendors in its first Magic Quadrant for this category, naming OpenAI, Anthropic, GitHub, and Cursor as Leaders and Tabnine the lone Visionary. The category is real money: Gartner sizes it at roughly $9.8 to $11 billion annualized as of April.

In the same report, Gartner cites its own 2026 Software Engineering Content Survey: 90% of engineering leaders report some productivity gains from AI, with a net average of 19.3%. That's the second number above, grounded in a measured survey result, already more conservative than either the vendor narrative or McKinsey's headline figure. Gartner's own forward-planning assumption goes further, projecting 30 to 50% productivity gains from asynchronous agent workflows by 2028, up from 0 to 20% for the code assistants of 2025. That's where Gartner expects the category to go, not where it is today.

The same report carries a second warning that rarely makes it into anyone's board deck. Gartner projects that by 2028, AI coding costs will overtake the average developer's salary, driven by rising token consumption and a shift from flat per-seat pricing toward consumption-based licensing. The productivity story and the cost story are sitting in the same document, and most of what gets repeated is only the first half.

McKinsey's 3 to 5x figure, published May 28, is the third number, and it's the one every vendor and every ambitious CTO wants a board to anchor on. It's real, for organizations that redesigned the entire delivery model around a twenty-four-hour human-agent handoff. It's also the largest number available, which is exactly why it travels the farthest.

A portfolio company CTO bringing a productivity story to the board is very likely quoting McKinsey's number, because it's the one built into the vendor's own marketing and the boldest consultancy research. The number that belongs in a value creation plan is Gartner's, at today's measured level, not the 2028 projection. METR's is the sanity check on how aggressive Gartner's number is allowed to get.

That gap matters because capacity gains do eventually show up in the multiple, if they're real, and if the cost of getting there doesn't eat the gain first. McKinsey's separate June study of 471 PE-backed companies found that those with the highest AI maturity traded at a median revenue multiple of 31x between 2023 and 2025, compared with 13x for companies using AI opportunistically. A capacity number that's really 19.3%, redeployed into roadmap and release velocity, is a genuine step on that ladder.

A value creation plan that counts the productivity gain and never revisits what the tooling itself will cost is pricing half the ledger. Gartner's own planning assumption is that the cost of tooling will overtake the average developer's salary within three years.

Price a value creation plan on the middle row. Treat the top row as a claim to verify, not a finding. Use the bottom row as the floor whenever a headcount number looks too good.

Three patterns have shown up consistently in conversations with AI engineering teams.

The first coding agent decision is a platform decision. Most teams don't find that out until they try to price a switch. Some of what gets built around an agent travels easily: instructions and rules files are plain text, cheap to keep in the repo, cheap to move. The harness underneath them, how the agent plans, verifies, and escalates, stays vendor-specific by design, and so does, however deep the workflow has grown into one vendor's repository, review, and CI integration. Keep the portable layer portable: plain markdown, versioned, and prefer agents built on MCP, the emerging standard for how an agent connects to tools and repositories, over proprietary integrations. The rest is a real cost. Price it like you'd price any other switching cost.

Writing the code has gotten a lot easier. Shipping hasn't caught up. Teams with real platform engineering and CI/CD automation already in place can integrate agent-generated code into the release pipeline with minimal friction. Teams without that muscle already built are finding that packaging, deployment, and release are still manual, and that's where the real bottleneck has moved. Security and compliance add a second brake on top of that: engineering wants broader repository and production access for these agents, and security and compliance decide how far that access goes.

The productivity story has a ceiling, and seniority sets it, for a specific reason: senior engineers carry context that an agent doesn't have. Business logic that was never written down anywhere. Edge cases nobody codified. That context is what lets them catch what an agent's output actually missed, not just how fast it arrived.

One case: an AI coding agent fixed a security issue across multiple repositories, and on the surface, the fix looked complete. A senior engineer reviewing the change caught that it handled roughly eighty percent of the actual cases and missed the rest, the kind of gap that only surfaces under review from someone who already understood what the fix needed to cover.

That's probably the same mechanism behind Gartner's own net average landing at 19.3% instead of near the 30 to 50% range projected for 2028, and behind METR's paradox too. The developers who felt 20% faster and measured 19% slower were likely doing the same thing: exercising judgment an agent can't see, which takes real time and never shows up as a keystroke saved. Junior engineers, working from a spec, ship what the spec says. Senior engineers ship what the spec should have said, and that's what real judgment costs in time.

Before approving any AI-driven headcount reduction, anchor the model on Gartner's own measured net average, 19.3% today, not McKinsey's 3 to 5x, and require the plan to name which bounded tasks are actually producing the number claimed.

Add coding agent governance, access controls, audit logs, approval policies, and data residency as a named line item in every technology diligence checklist. Treat it the way you'd treat SOC 2 or IP risk, not as an assumed detail inside an engineering interview.

Model the cost side of every AI coding agent plan with the same rigor as the productivity side. Gartner's own planning assumption has agent costs overtaking the average developer's salary by 2028, as pricing shifts from flat per-seat to consumption-based; a plan that never revisits that side of the spreadsheet is only half built.

19.3 percent. Gartner's own measured net average productivity gain from AI coding agents, from its 2026 Software Engineering Content Survey, cited in the same report that named four vendors Leaders. Ninety percent of engineering leaders report some gain. McKinsey's number for the furthest-along adopters is 60% smaller teams. Both are real. Only one is the number a typical portfolio company should be underwriting today.

Productivity improves. That much is true across every number in this issue. The actual diligence question is whether the assumptions behind the productivity figure, the implementation cost, the governance model, and the engineering workflow all describe the same company, or four different ones wearing the same board deck.

Portco Brief is a weekly briefing for PE operating partners and portfolio company executives focused on technology, AI, and value creation.

Subscribe at portcobrief.com.