The 70% your AI recognition vendor can't sell you

The promise attached to them is consistent. AI takes the operational burden. The human moment stays human. The gratitude still comes from a person.

Take that promise seriously, because in most cases it looks sincere. Look at where the AI in these products actually sits and you'll find it clustered in reporting, analytics, administration, and workflow. The components that come near the recognition message itself tend to coach rather than compose, prompting a manager toward specificity instead of writing the note for them. Several vendors have drawn that boundary at the feature level and published it. It's a defensible line, and more thought than the category usually gets credit for.

So grant the promise. Then notice the problem that no vendor selling AI recognition has any commercial reason to raise with you.

Their line doesn't protect you.

 The rule everyone is quoting is being quoted backwards

BCG's 10-20-70 principle turns up in nearly every serious AI strategy conversation right now. Ten percent of the value comes from the algorithm. Twenty percent from technology and data. Seventy percent from people and process.

The recognition category has adopted the number and inverted its meaning. It gets cited as though the 70% were a moral perimeter, the sacred human territory a responsible vendor agrees not to enter. That isn't what BCG measured. They were describing where the work is, and by extension where AI programs fail. The 70% is a labor estimate.

In a recognition program, that seventy percent has a fairly concrete inventory:

• Deciding which behaviors the program is actually trying to reinforce, and whether your managers can name them.
• Redesigning the moment recognition happens so it fits inside a manager's week instead of competing with it.
• Training people who have never been taught to observe work closely enough to describe it.
• Establishing who reviews coverage, how often, and what happens when a team shows up empty.
• Change management for a program most employees currently experience as a points balance.

None of that ships with a platform. When a vendor tells you the human part stays human, they're describing a constraint on their own product. They aren't telling you who performs the human part instead, on what schedule, with what training, accountable to whom.

That's your side of the contract. BCG's finding is that it determines roughly seventy percent of the outcome. 

What the other 30% is genuinely good for

This isn't an argument against AI in recognition, and a piece that landed there would be useless to anyone making a decision this year.

The 10% and the 20% are real and worth buying. Pattern detection is the clearest case. Recognition coverage is almost always uneven in ways nobody can see from inside the program, and a system that flags an employee who has gone eight weeks without acknowledgment is surfacing something a human manager reliably misses. Timing is another. So is the administrative load that keeps program owners buried in spreadsheets instead of in conversations with managers, which is the burden most of these launches are built to lift.

Those capabilities compound if the underlying process works. They compound in the wrong direction if it doesn't, which is the part worth dwelling on. 

Why this lands hardest in the mid-market

Picture the org this piece is written for. Two thousand employees. No dedicated L&D function. One person who owns recognition alongside four other programs, reporting to an HR Director who owns eleven.

Now ask the uncomfortable version of the question. Who on that team is doing the 70%?

Across organizations that have completed the Inspirus Recognition Maturity Model assessment, the average score is 18.4 out of 32, which falls in the Developing band. The average organization size in that sample is about 2,400 employees, which is to say the mid-market is the sample.

Developing is a diagnosis, not a euphemism. It means the program exists and functions inconsistently. Manager participation varies widely by department for reasons nobody has investigated. Coverage gaps go unnoticed until someone resigns and the exit interview surfaces them. The program has a budget and a platform and no owner with enough hours to run it as a system.

Layer AI onto that and you don't get a fixed program. You get a faster one. Better timing and better prompts applied to a process nobody has designed produce more of what already exists, delivered more efficiently, at greater scale. If managers are recognizing the wrong behaviors today, an AI that reminds them to recognize more often isn't an improvement.

The counterexample worth studying is what happens when the 70% actually gets staffed. ATCC, at 620 employees, reached 83% platform registration and reported a 10% improvement in retention in their first year on the platform. The registration number is the tell. Adoption at that level isn't something a product does to an organization. It's the output of program design, manager enablement, and sustained attention, which is the seventy percent doing its job. 

Three questions, and none of them are for the vendor

Most AI recognition checklists point outward at the vendor. Point these inward first, because they determine whether the vendor question even matters.

Who owns the 70%, by name? Not "HR," and not "the program owner." A person, with hours formally allocated to program design and manager enablement, separate from the hours they spend administering the platform. If the answer is the person already running recognition part-time, the implementation isn't funded, and adding AI doesn't change that arithmetic. It moves the bottleneck from administration to judgment. If you need to argue for those hours, the same logic that funds a platform funds the people running it, and our guide to making the business case for employee recognition applies to headcount as readily as to software.

What does a manager do differently on Monday? An AI can tell a manager that a direct report has gone eight weeks without recognition. Someone still has to have noticed something true and specific about that person's work in order to say anything worth reading. The notification creates the prompt. It doesn't create the observation, and it doesn't create the capacity to make one. If your managers can't describe a report's contribution in concrete terms today, what you're facing isn't a tooling constraint.

What breaks at volume? A manager with six reports and a light week can write something real. A manager with forty reports at quarter-end will take whatever shortcut is available, every time. That isn't a character failure, it's a load problem, and it's the exact moment your governance gets tested. Pilots get run by volunteers with time. Design the process for the manager who has neither. 

Then ask the vendor one question

Not whether their AI still feels human. Every vendor will pass that, and the ones that pass most fluently have simply spent the most on messaging.

Ask what they do to support the 70%. Manager coaching. Program design help during implementation and after it. Adoption benchmarking against organizations of comparable size. Someone who notices participation dropping in Q3 and calls you before you notice it yourself.

That question separates a platform from a partner, and it's the one most vendors are least prepared for, because the answer costs margin, scales badly, and can't be announced in a press release. 

Where we come down

Inspirus sells recognition software. Read all of the above with that conflict in view.

Our position is that this debate is being held at the wrong altitude. The interesting variable was never how human a vendor's AI feels, and the category has spent a year arguing about a question that the buyer's own implementation plan largely settles. What determines whether AI helps your recognition program is whether the seventy percent of the work that no product roadmap covers has a name, a budget, and a calendar attached to it.

Which makes this the right moment to disclose something. This fall we're releasing Connects AI Insights, and it belongs squarely in the 30%.

That's the honest description of it. It's a natural language reporting layer. Someone running a recognition program asks a question about their own data in plain language and gets an answer back, without building a report or waiting on an analyst. Participation trends, activity by team, recognition patterns across the organization.

The value is in what it makes visible rather than what it automates. A department can post strong participation numbers while one role inside it goes chronically unrecognized. A manager can believe they recognize their team regularly while the data says otherwise. Those are the gaps the tool is pointed at, and they're invisible from inside the program today.

One design decision is worth stating plainly. The intelligence layer and the data layer are segmented. The AI answers questions about your recognition activity without direct access to your underlying employee data, and your employee data is never used to train or operate the model. That constraint cost us capability. We think it's the right trade in a category asking HR teams to feed workforce data into models they can't inspect.

Connects AI Insights doesn't write recognition messages. It reports on them. Our published position on AI in recognition is that AI shouldn't generate appreciation on autopilot, and a reporting layer isn't a loophole in that.

We're naming which bucket it sits in because the alternative is to publish an argument about vendor candor and then quietly market against it. There's also an order of operations here, and we've said it elsewhere: consolidate first, measure second, add intelligence third. A reporting layer pointed at five disconnected sources produces a confident number that happens to be wrong. Pointed at a program you've actually staffed, it tells you faster than you can tell yourself which parts of the 70% you haven't.

The vendors drawing careful lines are drawing real ones, ours included. None of those lines does much for a program that's still Developing underneath it.

Before you evaluate anyone's AI, find out where your own program sits. The Recognition Maturity Model assessment scores the parts of the 70% that determine whether any of this works: manager enablement, frequency, visibility, alignment to values, and measurement. It takes about as long as a vendor demo and tells you something a demo can't.