The Last Six Percent

Written By Chantal Emmanuel

This week, the companies that sold the AI revolution started paying for what comes after the purchase. The bill tells you exactly where the real work lives.

Everyone was wrong about which part was hard.

The pitch for the last three years was that buying AI was the mountain. Pick the model, sign the contract, survive the integration, and the savings would arrive on their own. This week, in four separate stories, the companies that made that pitch said the quiet part out loud. The purchase was the easy 94 percent. The hard part is the last six.

I am borrowing that number from IBM, which found its AI could handle roughly 94 percent of routine HR requests. IBM is now tripling its US entry-level hiring, because the remaining 6 percent, the judgment calls, is where everything broke. Call it the last six percent: the slice of any job that looks like a rounding error until you automate everything around it and discover it was quietly holding the whole thing up.

The last six percent is the part of the flight an autopilot cannot fly. A modern autopilot handles almost the entire trip, and no airline fires its pilots because the plane can hold altitude on its own. The pilots are there for takeoff, for landing, for the weather nobody forecast, for the six percent that is all judgment and no cruise. For three years the industry sold the autopilot and let everyone assume it came without the cockpit. This week the bill for the cockpit came due.

You cannot automate the part you cannot define

Start with a person. Ford vice president of vehicle hardware engineering Charles Poon spent this week explaining why the company is bringing hundreds of experienced engineers back into roles that automated systems were supposed to cover. "Artificial intelligence is a fantastic tool, but it's only as good as the information you use to train it." Ford had bet that AI-driven quality control could catch design flaws. Without decades of engineering judgment encoded in its inputs, it could not. The tool ran fine. It just could not fly the six percent.

Ford is the loudest example, not the only one. Commonwealth Bank of Australia cut around 40 call center roles last year in favor of an AI voice bot, then reinstated them when call volumes rose instead of falling. Klarna, the Swedish payments company, ran the same cycle earlier: it claimed its AI assistant did the work of 700 agents, cut accordingly, watched satisfaction drop, and began rehiring. IBM automated its 94 percent and is now hiring into the 6 percent it could not.

The survey data puts numbers on the pattern, and the numbers are large. Orgvue, an HR analytics firm, found that 39 percent of business leaders made staff redundant because of AI deployment, and 55 percent of that group now admit those decisions were wrong. Robert Half found 32 percent of US hiring managers eliminated a role primarily because of AI and later rehired for the same or a similar position. Gartner projects that by 2027, half the companies that cut customer service headcount for AI will rehire for similar functions.

Two honest caveats, because the source matters. These figures come largely from HR vendors surveying their own market, so hold the exact percentages loosely. And the financial case for reversal is mixed rather than damning: Careerminds found nearly 31 percent of companies spent more on rehiring than the original cuts saved, and another 42 percent merely broke even.

One thing should not get lost inside the business analysis. Every percentage point in those surveys is a person who lost work while an employer tested a theory. Some of those roles are coming back, sometimes at different pay, sometimes filled by different people. The reversal is a correction, not a happy ending for everyone who lived through it.

Setup is now the product

If the rehiring wave shows what the last six percent costs when you ignore it, this week two of the largest companies in the world put a price on getting it right.

On Thursday, Microsoft announced Microsoft Frontier Company, a new operating business backed by 2.5 billion dollars and 6,000 industry and engineering experts. Its mission is not building new AI. It is deploying existing AI successfully inside customer organizations: evaluating models, integrating them into business processes, building the implementation strategy. Two days earlier, Amazon Web Services committed 1 billion dollars to its own deployment venture. Earlier this year, the same embedded-engineers model helped propel Palantir's growth, and now the two largest cloud companies on earth are copying it at scale.

Read that the way I do. The sellers of AI just put a market price on the gap between buying AI and benefiting from it, and the price is billions. That tracks with everything the enterprise data has shown all year. Rising usage costs, unpredictable budgets, stalled pilots: these are not model problems. They are workflow problems. Unclear processes, bad data, no human checkpoint, no one who owns the outcome. The six percent, again, wearing a different suit.

Here is the practitioner read, and it is the same thing I tell BAMPT clients in week one. The tool is step one. Results come from mapping the workflow before automating it, deciding where human judgment stays, training the people who will run it, and measuring cost per completed task instead of excitement per demo. Microsoft just validated that entire checklist with a 2.5 billion dollar business unit. You do not need 6,000 experts. You need one honest audit of one workflow.

When the six percent gets a seat in Washington

The third story raises the stakes on the first two, because it says the thing everyone now depends on has grown too important to sit outside the government's reach.

OpenAI, the company behind ChatGPT, has proposed giving the US government a 5 percent ownership stake, worth roughly 42.6 billion dollars against the 852 billion dollar valuation from its March funding round. The Financial Times broke the story Thursday, and the detail makes it bigger than one company: OpenAI CEO Sam Altman suggested every leading US AI developer contribute the same 5 percent to a fund modeled on the Alaska Permanent Fund, which pays annual oil dividends to Alaska residents.

The caveats are heavy. The talks are described as conceptual and early-stage. Neither OpenAI nor the White House has confirmed anything, an actual deal might require an act of Congress, and no rival lab has signaled it would participate. Reuters reported that at least one competitor has had no such discussions with the administration at all.

So why does a maybe-deal clear the bar? Because of the precedent stack underneath it. The federal government already holds a stake in Intel, converted from chip manufacturing grants. Chipmakers have agreed to revenue-sharing arrangements on certain AI chip sales. A June executive order created a framework for government review of frontier AI models before release, and one major model spent most of June offline under export controls. Piece by piece, without much fanfare, the relationship between the government and the AI companies is shifting from regulator-at-a-distance to something more entangled.

I am not going to tell you whether that is good policy. That is not my lane. What I will tell you is what it means operationally: the tools your business increasingly depends on are being negotiated at a national level, the way energy and telecom are. That affects long-term pricing, access, and stability. When your vendor's availability can change because of a conversation in Washington, vendor concentration risk stops being a hypothetical and becomes a line item.

What this actually means

Three stories, one direction of travel. The rehiring wave says judgment cannot be cut out of the loop. The deployment businesses say setup is now the product. The ownership talks say the stakes are large enough that governments want a seat at the table. Put them together and the week's real headline is this: the industry has moved past "having AI" as an advantage. The companies winning the next phase, at every size, are the ones that run it well. It is a people-and-process story wearing a technology costume, and it has been all along.

So here is what to do with that, in plain terms.

Before you cut or reassign any role because of AI, separate the tasks from the judgment. The surveys above are full of leaders who automated the tasks, lost the judgment, and paid twice to get it back.

If you are planning AI spending for the second half of the year, budget for implementation, not just licenses. As a rough posture, expect the setup work, the process mapping, the data cleanup, the training, the checkpoints, to cost as much attention as the tool itself.

Run one workflow audit this month. Pick a repeated weekly process, map it in plain language, test AI on it with real material, and add a human checkpoint before anything reaches a customer.

Note your vendor concentration. If one AI provider going offline for three weeks would hurt your operations, write down your fallback now, while it is a planning exercise and not an emergency.

And if you already made AI-related cuts and results have slipped, you are in the majority, not the minority. Reintroduce oversight deliberately, with a defined role, rather than quietly re-adding headcount and hoping no one asks.

The autopilot has been on for three years. This week, four different companies, a carmaker, a bank, the two biggest clouds, and the most valuable AI startup in the world, all remembered the same thing at once. The cruise was never the hard part. The hard part is the six percent the machine hands back to you right when it matters, and whether there is still someone in the seat who knows how to fly it.

Chantal Emmanuel is the co-founder of BAMPT, where she builds AI automation systems for service businesses, and the CTO of LimeLoop. She covers AI news for business owners every week.

SOURCES

Chantal Emmanuel
Next

Optimizing Marketing Funnels. The Pause Is Doing More Work Than the Data