How to Adopt AI Without Deskilling Your Team

DOWNLOAD NEW WORKSHEET: Four practices that keep AI from deskilling your team

In 2025, The Lancet published a study of 19 experienced endoscopists in Poland, each with thousands of procedures behind them. After they began using AI-assisted detection, their unassisted detection rate for precancerous growths fell from 28.4 percent to 22.4 percent. About a fifth of their skill lost, in months, among acknowledged experts.

The same pattern is arriving in our organizations, quietly, in grant narratives nobody drafts cold anymore and community surveys nobody reads in full.

This is not an argument against the tools. We use them, we recommend them, and organizations that ignore them will fall behind organizations that do not. It is an argument for four practices that cost almost nothing and protect the thing that makes the tools worth having in the first place.

1. Write the question before you open the tool

One sentence, on paper, before the chatbot loads. What am I actually trying to figure out?

The teams we work with that do this consistently get noticeably better output, for an unglamorous reason. The human did the part the machine cannot do, which is deciding what the problem is. A vague prompt produces fluent text that answers a question nobody asked.

It also leaves a record. Six months of those sentences tells you what your organization actually uses AI for, as opposed to what your AI policy says it uses AI for. Most leadership teams have never seen that gap measured.

2. Schedule unassisted reps on purpose

Decide which capabilities your organization cannot afford to lose. For most of us it is a short list: writing a case for support, understanding a community, reading a budget, running a hard meeting, sitting with a donor who is upset.

Then practice them without assistance, deliberately, on a schedule. Draft the appeal cold once a quarter. Run one meeting a month without the summary bot. Have someone senior review a grant narrative before it goes anywhere near a model.

Microsoft’s own 2026 research found that its most advanced AI users do exactly this, at 43 percent against 30 percent for everyone else. The people deepest into these tools are the ones protecting their skills hardest. That is worth sitting with.

3. Adopt a sourcing rule

Nothing leaves the building unless every factual claim traces to a source a named human checked. Board reports, grant applications, donor communications, program evaluations, the statistic in the annual appeal.

Fabricated statistics now ship in seconds, and they arrive fluent, confident, and formatted exactly like real ones. One of them in front of the wrong funder costs more than every hour the tools have saved you, and the damage lands on your credibility rather than on the vendor’s.

Make it a named role rather than a general expectation. General expectations are how this fails.

4. Keep a register and put a name on every use

One page. Which systems you use, which decisions they touch, who owns each one, and what you will not use them for. Review it quarterly at the leadership level, not annually at the staff level.

The register matters most for the decisions being delegated quietly, without anyone deciding to delegate them: screening applicants, scoring donors by likelihood to give, triaging clients by need. Those are exactly the decisions a board and senior staff are supposed to own, and they are the easiest ones to hand to a model by default because the model is fast and the queue is long.

Regulators are moving in this direction and will arrive eventually. The better reason to do it now is that you cannot govern what nobody has written down.

The part that is easy to miss

AI hands your people hours back. The instinct will be to convert those hours into more output, because scarcity trained us to convert everything into more output.

Convert them into field time instead. Donor visits, community conversations, the listening session where somebody says the thing that never makes it into a report. That unwritten knowledge is the only durable advantage our organizations hold over a tool that anyone can buy for $20 a month.

Spend the dividend where the machine cannot follow.

This post is the practical companion to Issue #7 of The Social Prophet, on the deskilling evidence and the human capabilities that are appreciating in value. Read it at thesocialprophet.org.

Common Ground Consulting helps nonprofit and foundation leaders adopt AI with the governance to make it stick. If you are building this out, get in touch.