The Signal Issue 6.

Two numbers from McKinsey have been quoted in every AI deck since November. 88% of organizations regularly use AI in at least one business function. 39% attribute any level of EBIT impact to it.

Note what the first one actually measures. At least one business function. That is the lowest bar a survey can set, and nearly nine in ten clear it. On adoption, this is over. There is no holdout population left to convert.

Note what the second one actually measures. Any level of impact. Not a large impact. Any. Around 6% of organizations attribute more than 5% of EBIT to AI, and nearly two-thirds say they have not yet begun scaling it across the enterprise.

That report is nine months old. Everyone can recite the gap. Almost nobody has explained it, and a statistic that gets repeated for nine months without a mechanism stops being an insight and becomes a piece of furniture.

This week two publications moved it. One says why the gap exists. The other shows what closing it looks like from the inside.

Neither of them is about a model.

Which is why Grab’s engineering post is the most useful thing published this week, despite carrying the smallest numbers in the issue.

Grab measured the share of tickets its data analysts closed that were mechanical: data preparation, alerting, routine reporting. In February 2026 that was 44%. By June 2026 it was 30%. Over the spring, the share of requests answered with no human involvement rose across the board, with SQL requests going from 50% to 81%, data pulls from 63% to 90%, and metric questions from 53% to 67%. Cycle times fell by roughly a third. The freed capacity was redirected to building self-serve workflows for stakeholders and to deeper analysis.

Take the caveats first, because they are large. This is Grab’s own engineering blog. One function, one company, no independent audit, no EBIT figure anywhere in it. It is a documented internal result and it is not a benchmark. Anyone quoting 44 to 30 as what companies achieve is doing the thing this brief exists to argue against.

Now take what it establishes, which no survey in this issue can.

Grab knew that 44% of that work was mechanical before it changed anything.

That is the rare part, and it is not the improvement. It is the baseline. A company that has never counted the mechanical share of its own work can install the same agents, get the same result, and report precisely what the 39% reports: nothing measurable, because nothing was measuring. The investment is visible because procurement recorded it. The return is invisible because no ledger exists for time that stopped being wasted.

There is a second detail worth borrowing. The recovered capacity was pointed somewhere specific. That is not a soft point about morale. Time recovered from coordination does not sit still waiting to be counted. If nobody decides where it goes, the organization reabsorbs it, and by the next quarter there is nothing to show a CFO. A real share of honest “we saw no impact” reports are companies that genuinely recovered time and never assigned it.

Three limits, stated plainly.

First, none of this is a random sample. McKinsey surveys its own respondent pool. Deloitte surveys organizations already piloting agentic AI, which selects for the committed and makes its readiness figures a best case rather than an average. Grab reports on itself. The convergence across them is the signal. No single figure here carries an industry.

Second, one number that would have fit this issue neatly has been left out. A survey circulated this week reporting that one in three organizations already use AI in critical resilience workflows while 30% have never tested for agentic failure. It supports the argument in this brief almost too well. Its sample size and methodology are not disclosed anywhere in the coverage, and the underlying report could not be found. A percentage with no denominator is a rhetorical device. It is not in this issue, and the reason is printed here rather than buried.

Third, the governance evidence that is solid points somewhere uncomfortable. Check Point researchers spent about a year attacking the frameworks enterprises build agents on and disclosed 11 vulnerabilities across LangChain, LangGraph, CrewAI, AutoGen, Microsoft Agent Framework, and Google ADK. These are not prompt-injection curiosities. They are old, boring bug classes, deserialization, server-side request forgery, path traversal, in new agentic plumbing. Microsoft patched a checkpoint-deserialization flaw that allowed remote code execution and paid a bounty for it.

That belongs in an issue arguing the constraint is organizational, because it is the counterweight. Some of the failure to scale is caution, and some of that caution is correct. A brief that only prints the numbers agreeing with its thesis is an advertisement.

Not 88%, which is finished and tells you nothing you can act on. Not the $7 billion, which is somebody else’s conviction priced in somebody else’s currency.

16% is the share of leaders who say their business processes are prepared for agentic adoption, in a survey of organizations already piloting agents. The same people put 74% on the expectation that nearly half those processes get rebuilt within four years.

Seventy-four percent see the work coming. Sixteen percent are ready to do it. That distance is not a vendor problem and it is not a budget problem. It is the reason nine-in-ten adoption produced four-in-ten profit impact, and it is the only number on this page you can change from inside your own building.

Start by finding out where you actually sit. readiness.align-ify.com. 33 questions, free, ends with a real number.