You Cannot AI Your Way Out of a Mess You Do Not Understand
Technical debt was the number one Salesforce admin challenge for the second year running, and only 2% call their org clean. Agents turn that from a slow tax into an invoice. Here is the rubric I score orgs on before anyone grounds an agent in them.
Every Salesforce team I talk to right now is trying to bolt an agent onto their org. Leadership saw the keynote. There is a mandate. The clock is ticking.
Almost nobody is asking the question that comes first: what happens when you point AI at an org that no one fully understands any more?
The uncomfortable answer is that AI does not fix a mess. It inherits one. Grounded in clean, well modelled metadata, an agent is genuinely useful. Grounded in ten years of overlapping Flows, dead fields and automation nobody documented, it inherits every bit of that tangle, and now it acts on it, at speed, in front of your customers.
The hype cycle sells building faster. The thing quietly breaking teams is that they cannot safely change what they already have.
The bill is coming due
This is not a hunch. In the 2026 State of the Salesforce Admin survey, technical debt was the number one challenge for the second year running, named by 56% of admins. Only 2% described their org as clean and well maintained. Two percent.
Meanwhile the delivery gap is real. Despite AI being used almost everywhere, only about a third of Salesforce developers have actually shipped an AI project. The distance between using AI and shipping with it is mostly the mess in between.
For years, technical debt was a slow tax. You could defer it. AI is what turned it into an invoice, because AI works badly on top of a messy org and it does so loudly.
The scarce skill is judgment now
When AI can generate the Apex, build the Flow and draft the agent, generation stops being the constraint. What becomes scarce is judgment: knowing which problems are worth solving, which changes are safe to make, and which requests to decline.
That is not a tooling problem, and it is worth being clear about why nobody is going to sell it to you. Salesforce will not solve it, because the platform is where the debt accumulates. Your DevOps vendor will not either, because a pipeline moves changes, it does not tell you whether a change is wise. This is the work of understanding the org you actually have, as opposed to the org you think you have.
Clarity before agents. Always in that order.
A way to see it: the AI-Readiness Index
Here is the rubric I use to make org clarity concrete instead of a feeling. Score each dimension from 0 to 5, where 0 means nobody knows, 2 means it is known and unmanaged, 4 means documented and controlled, and 5 means clean. The rollup is the index: a blunt read on whether you are ready to let agents loose, or about to automate a mess.
- Automation sprawl.Overlapping Flows, and the Process Builders and Workflow Rules still running underneath them with no rule about which one wins.
- Dependency risk.What breaks if you change one field. Hardcoded IDs, and integrations nobody wants to touch.
- Documentation coverage.Undocumented logic, tribal knowledge, and what happens the week the person who built it leaves.
- Permissions hygiene.View All and Modify All sprawl, and profiles nobody has audited since they were cloned.
- Dead metadata.Unused fields, obsolete classes, and components still pinned to a retiring API version.
- Data quality.Duplicates, half filled records, and free text that no model can reason over reliably.
- Change safety.Real test coverage, a rollback that has been rehearsed, and sandbox discipline.
- AI-groundability.The honest one. Is the metadata clean enough for an agent to be grounded in good data?
A high score does not mean your org is impressive. It means it is legible: to a new admin, to your future self, and now to a model.
You can score yours on theAI-Readiness Index, which does the arithmetic and tells you which two dimensions to start on.
Where to start
Do not try to fix all eight. Score them, take your two lowest, and start there.
It is usually automation sprawl and documentation, because those two compound into everything else. Undocumented, overlapping automation is exactly what makes dependency risk unknowable and change unsafe. Fix those two and the other six get cheaper.
The teams that win the next two years will not be the ones who adopted AI fastest. They will be the ones who understood their own systems well enough to trust what the AI did with them.
Clarity before agents.
What is the worst piece of technical debt hiding in your org right now? Tell me. I am collecting the ugliest examples for a teardown series, names filed off.
