Gartner expects agentic AI to resolve eight in ten common service issues by 2029. Every vendor now quotes it, including us. It is worth reading carefully, because the word doing the work is resolve, and resolution is not a model capability.
Resolution is an integration problem
Take the ten most common requests in any contact centre. Where is my order. Cancel this. Change my appointment. Send the invoice. Is it under warranty. Reset it. Where is my result. Every one of them is answered by a system you already own, and none of them is answered by a language model. The model decides what the customer wants; your order system decides what is true.
Which means the path to eighty percent runs through integration coverage, not model quality. An organisation with two systems connected will plateau early with the best model available. An organisation with eight connected will get there with a modest one.
The second half: permission
Reading from a system is a technical decision. Writing to one is a governance decision, and it is where most programmes quietly stop. Who signs off that the AI may issue a refund under a certain value? What is the approval boundary? What happens on an exception? Those answers come from your operations and risk functions, not from procurement, and they take longer to obtain than the integration takes to build.
What to do with the forecast
- Rank your top twenty request types by volume, then mark which system holds the answer. That list is your roadmap, in order.
- For each, decide read-only or write, and get the write decision signed before you build.
- Measure completion, not deflection. Deflection improves when the system is unhelpful enough that people give up.
- Assume the last twenty percent is not coming. Those are the cases that need judgment, and they are what your people should be doing.
The forecast is not a prediction about models getting better. It is a prediction about enterprises finally connecting their systems. The organisations that hit it will be the ones that treated it as an integration and governance programme with an AI component, rather than an AI programme with an integration annoyance.