Site icon Kevin Waites Consulting Services

How everyday corrections train your AI assistant

A lot of nonprofit tech projects fade out slowly. I call it the maintenance tax. Each week someone spends a bit more time fixing the tool and a bit less time using it, until the team quietly goes back to the old spreadsheet.

AI can go the same way. You ask for a social post, the tone’s off, so you fix it and move on. Next week you fix the same thing again. A month in, you’re wondering whether the tool saves you any time at all.

With an AI agent that keeps a memory, those fixes can pay you back. Explain a correction once and the agent can carry it into every task after that. The time you spend correcting turns into training, and the tool gets more useful the longer you work with it.

Why the explanation matters

Most people fix AI output quietly. They change the line and get on with their day. The AI never finds out what was wrong, so it makes the same choice next time.

An agent learns a lot more when you tell it what you changed and why. Here’s the kind of note I give mine:

“I changed the ending of this post because we say ‘mission’ instead of ‘charity.’ Use ‘mission’ in every future post.”

That one sentence gives the agent a rule and tells it where the rule applies. The reason helps too. Once the agent knows why you made the change, it has a better shot at handling similar cases you never mentioned.

The same goes for facts. If you fix a number or a program name, tell the agent where the right version lives, like your annual report or your program list, so it checks there next time.

Narrate as you work

The other half is narration. You don’t need to write a 50-page manual to train an AI assistant. Tell it what you’re doing as you move between tasks, the way you’d bring a new hire along:

When it makes suggestions, give it quick feedback. “This part was helpful, that part was too formal” is plenty.

Over a single week, that loop takes you from generic output to drafts that sound like they came from someone who works with you.

What stays when people leave

In most organizations, when a staff member leaves, the lessons they learned go with them. The next person learns the house style and the funder quirks all over again, usually by getting them wrong first.

When corrections are saved in the system, they stay put. A new hire drafting their first social post gets the same rules everyone else follows. Each mistake someone takes the time to explain makes the system a little sturdier.

How to make your corrections count

  1. Stop fixing quietly. When you change something the AI wrote, take ten seconds to tell it what you changed and why.
  2. Say how far the rule reaches. “For this email” and “for every donor email from now on” are different instructions. Spell out which one you mean.
  3. Set the scene before you start. One or two sentences on what you’re working on and what you care about gives the AI a head start on the first draft.
  4. Keep your feedback short. “Helpful,” “too formal” or “wrong program name” is enough for it to adjust. You don’t need to write a paragraph.
  5. Put team-wide rules where the team can use them. If your tool has shared or project-level instructions, save organization-wide rules there so everyone’s drafts follow them.

For leaders: count corrections as training time

Your staff will probably read every correction as a sign the AI isn’t working. Help them see it differently. A ten-second explanation today saves the same fix next week, and the week after that.

It also changes how you judge an AI pilot. A tool that needs a lot of corrections in its first few weeks can be fine, as long as the corrections stick. Worry about the one that needs the same fix in week six that it needed in week one.

Questions worth asking your team:

Learning from each correction is one of the main ideas behind Cosly, the AI staff team I launched last week for small nonprofits.

Exit mobile version