AI grafting is what happens when a company attaches AI to an existing process and calls that transformation.
The old workflow remains intact. Same intake form. Same queue. Same approvals. Same handoffs. Same meeting. Same manager review. Same definition of done. The only visible change is that one step now has an AI button beside it.
This is tempting because it is easy. Nobody has to renegotiate ownership. Nobody has to change the process diagram. Nobody has to decide which work should disappear. You add AI where the work already exists, capture a before-and-after time saving, and declare progress.
The problem is that AI grafting preserves the assumptions that made the workflow slow in the first place.
If the old process had too many handoffs, AI makes each handoff receive more material. If the old process had weak quality standards, AI produces more outputs that nobody can judge consistently. If the old process had unclear decision rights, AI increases the number of cases where someone must decide whether the output is good enough. If the old process hid work in Slack, spreadsheets, and manager memory, AI depends on context that the system still cannot provide.
The workflow looks modern. The operating model is unchanged.
You can spot AI grafting by looking for a few symptoms.
First, the process map has not changed. A team says AI transformed the work, but every box and arrow is the same. The only difference is that one box now says "AI-assisted." That usually means the team automated labor inside the box without asking whether the box should still exist.
Second, review demand rises. AI produces drafts, recommendations, summaries, and classifications faster than humans can absorb them. The team celebrates production speed while managers quietly become the new bottleneck.
Third, exceptions increase. The happy path got faster, but edge cases now require more explanation because nobody redesigned escalation rules. People are left asking, "Is this one safe to trust?" several times a day.
Fourth, accountability gets fuzzy. The person using the AI feels like the system produced the work. The reviewer feels like they are only checking. The manager assumes the owner has it covered. The customer experiences the result as one company, not a chain of partial responsibility.
Fifth, the old meeting survives. If the workflow were truly redesigned, some meetings would change or disappear. When the same review meeting persists with more AI-generated material in the packet, the company probably grafted output onto old decision habits.
Grafting is not always useless. It can be a reasonable first experiment. A team may need to learn where AI helps before redesigning the workflow. The mistake is treating the graft as the end state.
A better sequence is simple.
Start by identifying the work outcome, not the task. For example: resolve a customer issue, qualify a lead, close the books, prepare a renewal, ship a release, approve a contract, investigate an incident.
Then map the current work unit. What starts it? Who owns it? What context is needed? Where does it wait? Who reviews it? What decisions happen? What exceptions appear? Where does quality fail?
Only then decide where AI belongs. Some steps should be assisted. Some should be reviewed by AI before a human sees them. Some should become exception-only flows. Some can be automated completely. Some should remain human because the value is judgment, trust, negotiation, or accountability.
The redesign often removes work rather than speeding it up. Maybe the AI drafts the customer note, but the bigger win is eliminating a handoff between support and product because the issue pattern is now automatically summarized into a product queue. Maybe the AI helps with renewal prep, but the real change is that account risk signals are reviewed weekly instead of recreated before every QBR. Maybe the AI speeds legal review, but the work redesign is a clearer contract playbook and exception path.
That is the difference between grafting and redesign.
Grafting asks, "Where can we add AI to this process?"
Redesign asks, "What should this work become now that AI changes production, review, and exception handling?"
The second question is harder. It creates arguments about ownership, quality, control, and role boundaries. Good. Those arguments are the work. If a company avoids them, it gets a shinier version of the old system.
This is part 2 of 10 in Work Design for the AI Era.
