Jason LeeLattice

Item 51 of 55

Lattice tried to put a bot on the org chart, and the bot fit

5 min 1,234 words

A printed employee performance evaluation form with rating boxes, the paper ancestor of the employee record

The calibration meeting I remember best started at four in the afternoon and ran until a few minutes past ten, and the whiteboard had been gridded before anyone arrived: nine boxes, three rows, one row per rating the firm was willing to hand out that cycle. A partner moved a name from the middle row to the top one, saying the associate had stepped up on a matter whose details he was reconstructing as he spoke, and the name stayed where he put it. Eleven minutes a person. I know the number because an analyst kept time without anyone asking him to, the way people in that profession keep time, quietly and against their own interests. Somewhere in the fourth hour a name moved back down, and with it went a year of work: the two in the morning rebuilds, the recovery of a client who had gone cold in March, the quarter the associate spent as the only person on the account who understood the model. None of it was discussed. It became a position on a curve, and the curve had a quota, because the firm had read about distributions and decided it should have some. Nobody in that room disliked the associate. That part matters. The room was not cruel; it was administrative.

In July 2024, Lattice, a company that sells people-management software to roughly 5,000 customers, announced that it would give AI "digital workers" employee records, onboarding, managers, and performance reviews inside its platform. Three days later it withdrew the plan. The announcement had been received online as something between a novelty and a violation, and the withdrawal was reported with the tone of a company backing out of an embarrassing commitment. Roy Maurer covered the reversal for SHRM here. The mockery was close to unanimous, and it was aimed at the wrong thing. The announcement wasn't tone-deaf. It was candid. A person inside Lattice is already a row with a manager field, goals attached, a pulse survey, a compensation number, and a rating that arrives on a schedule, and the software did not need to be taught to see employees as records, because a record is the shape of the object it stores. The bot fit the way a filler fits a form. The form was drawn for it. A performance suite that can imagine a digital worker hasn't misunderstood what a person is. It has finally said, in public, what it believed a person was all along.

A row does not know it is a row

Lattice's modules sound like a curriculum, but they are views onto one object. Performance is the record's rating. Goals and OKRs are its attachments. Engagement is its answers. Compensation is its number. Grow is its trajectory. When the company proposed digital workers, the engineering required was almost nothing, because an AI agent already has everything the schema asks for: a name, an owner, tasks that can be written as goals, output that can be scored. That's the interesting part. The product work was never the obstacle; the obstacle was that the announcement read the schema out loud, in a keynote, with the word "employee" attached. Two years on, the vocabulary has moved and the row has not. In June 2026 Lattice launched Workforce Intelligence in early access, framed as connecting people strategy to business outcomes, which is the same record wearing a more expensive jacket. The row is the unit of sale, the unit of analysis, and, in the company's imagination, the unit that anything can occupy. A row does not know it is a row. That was always true of the human ones too.

Follow the money before following the outrage. The buyer of this software is HR, and behind HR is legal, and what each of them purchases is not insight. The most durable output of a review cycle is documentation, and documentation is written for the reader who arrives later, which is usually a lawyer. Lattice prices by seat, per module, per month, annual billing only, with a $4,000 floor per account, and the modules stack, per the vendor's own pricing page at the start of October 2026:

Module Per seat per month
Performance $10
Goals and OKRs $8
Compensation $6
Engagement $4
Grow $4

A five-module stack is $32 per head per month, and this structure rewards exactly one behaviour: attaching more modules to more heads. Nothing in it pays for a good conversation, and nothing could, because a good conversation leaves no artifact. The company has about 5,000 customers, a valuation Fortune put at $3 billion in 2025, and a CEO, Sarah Franklin, who arrived from Salesforce rather than from a founder's garage. This is a mature vendor selling records to a buyer whose own exposure is unrecorded liability, and the digital-worker episode reads clearly inside that arrangement. The vendor wanted a record for the bot because a record is what it knows how to sell. Consider the incentive. The scandal was never that Lattice disrespected human beings. It was that a record has no field for respect, and respect therefore cannot be billed.

The record protects people, and that is worth conceding

The counterargument deserves better than the mockery gave it. Structured, documented reviews protect employees. The alternative to calibration software is not a purer conversation between honest people. It is an unrecorded conversation among people who like each other, conducted behind a closed door, and in that arrangement the associate's year depends entirely on whether his partner ate lunch. The strongest evidence for this pessimism about human judgement comes from the other side's own literature, which is what makes it useful. Scullen, Mount and Goff analysed two large datasets of performance ratings for the Journal of Applied Psychology in 2000 and found that idiosyncratic rater effects, the portion of a rating explained by which rater you drew rather than how you performed, accounted for 62 and 53 percent of the variance across the two datasets. More than half of a performance rating measures the rater. Both things are true: a documented curve with a quota is more honest than a manager's confident memory, because the curve at least admits it is a curve, and the record is then mostly fiction, made durable and citable. I have sat on both sides of this. In consulting I was a position on a whiteboard. At the software company I was a row in a system like this one, reviewing and reviewed, and I would take the row. The row can be read back. The whiteboard could not.

The verdict depends on a variable easier to name than to see. If the row stays a document about human beings, Lattice is the least bad version of something that has to exist, and the retreat from digital workers was a company learning what its customers will let it say out loud. If the row becomes a table anything can occupy, then the humans are the entries that need reading carefully, and that day will not be announced, because a schema change never is. Somewhere there is a whiteboard with a bot's name in one of the nine boxes, moved in eleven minutes by a partner who never wondered what the bot thought of him. The bot will not know it was scored. Neither, in the way that mattered, did the associate.