Aisha Amaira is a powerhouse in the MarTech world, specifically when it comes to the messy intersection of CRM architecture and revenue operations. With a deep background in customer data platforms, she has made a career out of untangling the digital knots that stop sales teams in their tracks. We’re sitting down to talk about why even the most sophisticated AI-assisted CRM updates can fail if the underlying logic is buried under years of technical debt. We dive into the specific challenges of managing hundreds of thousands of records, the dangers of conflicting data enrichment tools, and the rigorous auditing process required to ensure AI agents act as a scalpel rather than a sledgehammer.
Many B2B firms struggle with hundreds of thousands of records where only a tiny fraction meet basic quality standards. How do you approach the initial chaos of a CRM that is actively halting sales operations?
The reality on the ground is often staggering because research shows only 3% of companies have data that meets basic quality standards. I recently walked into a B2B software firm that was a textbook example of this; they had hundreds of thousands of records, including customers, deals, and a mountain of inactive or duplicate data. This level of poor data quality isn’t just an eyesore for the RevOps team; it costs organizations an average of $12.9 million a year. In this specific case, the chaos was so intense that the sales team couldn’t even trust their lead assignments because the system was showing contacts in the wrong regions entirely. I had to start by recognizing that the “system” wasn’t just broken, it was actively fighting the sales reps, leading to a constant stream of complaints every time a lead landed in the wrong territory.
You mentioned that an AI-assisted update can result in zero failed writes but still leave the data incorrect. What is happening behind the scenes that allows for this kind of “successful” failure?
It is a deceptive feeling of victory when you see a “0 failed writes” notification because it simply means the database accepted the change, not that the change was correct. In a recent project, I processed 42,185 changes in about 8 minutes, and while the software confirmed every write was saved, the underlying logic was being sabotaged by old workflows. I found that even though I had written the correct data, an old routing workflow—one that we hadn’t even realized was active—triggered just minutes later and reverted two of those records back to the old, wrong owner. This happens because people often build new workflows on top of existing ones without turning the old ones off, creating a literal crash of logic. You can have a perfect AI agent execute a plan, but if you haven’t audited the existing 327 live workflows, the system will just “successfully” write the wrong information over and over.
When dealing with three different enrichment tools providing contradictory data, how do you establish a “single source of truth” without causing even more system conflicts?
The conflict usually stems from the fact that these tools—whether they are third-party apps or HubSpot’s native enrichment—start communicating with each other and overwriting data without any hierarchy or rules. In the company I worked with, they were using three different tools that were essentially shouting over each other, leading to records flipping regions constantly. To fix this, I had to pause 14 of the 17 workflows that had the power to change company owners and force a hard stop on the “automated” chaos. I had to manually decide the conditional logic rather than letting the AI guess, specifically looking at things like whether a rep had an open deal with an account. If a rep is already talking to a customer outside their region, it makes no sense to let an enrichment tool’s regional tag override that human relationship, so I kept those records out of the change list entirely.
Using AI agents to prepare work that would normally take months is a massive efficiency gain, but how do you manage the risk of the AI making those 42,185 changes too quickly to catch a mistake?
The speed is definitely intoxicating—doing weeks of work in 8 minutes is incredible—but you have to treat the AI like a powerful engine that needs very strong brakes. I used Claude Code to read the workflow definitions and build the change lists, but I never allowed the AI to actually hit the “send” button on the final logic. I built the change lists into a fixed list first, and then used a separate, plain script to make the writes in batches of 100 records at a time. This batching allowed me to monitor the progress and ensured that if something went sideways, I wasn’t blowing up 55,000 accounts in a single second. Most importantly, I ran the entire process on a sandbox copy of the portal for several rounds, which is where you catch the sensory details of the system—like seeing a field clear when it should have stayed populated—before touching the live production environment.
What specific guardrails do you put in place to ensure that protected accounts, like those with open deals, aren’t accidentally reassigned during a mass cleanup?
Guardrails are the only thing standing between a clean CRM and a total operational disaster, especially since Gartner predicts that by 2027, 40% of enterprises will decommission AI agents due to governance gaps. For my run, any company marked as a target account, tagged as a partner, or on a do-not-sell list was strictly off-limits for reassignment, regardless of what the regional data said. I even performed a manual check of 992 open deal accounts after the run and discovered that one had actually moved because a deal was created seconds after my initial data pull. Because I had a secondary script re-reading every single record after the write, I caught that one outlier and restored it the very same day. You have to anticipate that the system is “breathing” and changing even while you are working on it, so you need that final audit to compare the results against your original plan.
If you were to approach this same project again, knowing how those two accounts were reverted by an old workflow, how would your audit process evolve?
I would be much more aggressive in my “pre-flight” check by re-reading the status of protected accounts literally seconds before the writing process begins. My first audit identified 17 workflows that could write to the owner, but I only paused 14 of them because we assumed the others were harmless; that was a mistake because those remaining workflows still had the power to move data back. Next time, I will check every single integration, enrichment rule, and even manual user permissions that could possibly touch a record. We have to stop accepting “the update succeeded” as a final answer and instead ask, “What else in this ecosystem has the permission to change this record five minutes from now?” It’s about moving from a mindset of a “one-time fix” to a “continuous verification” model where you are always suspicious of the system’s old habits.
What is your forecast for how AI-driven RevOps will change the role of the CRM administrator over the next couple of years?
By 2027, I expect the role of the CRM administrator will shift from being a “data entry gatekeeper” to becoming a “logic architect” who spends more time auditing AI agents than manually fixing records. We are moving toward a world where 40,000 changes can happen in the time it takes to grab a cup of coffee, which means the margin for error is shrinking to near zero. Administrators will need to become experts in “reading” code and workflow definitions via AI to find the “hidden” triggers that cause data decay. If you don’t understand the conditional logic underlying your automation, the AI will just help you make mistakes at a much faster, much more expensive scale. My advice for readers is to never trust the software’s “success” message; always build a second script to re-read your data and prove that the plan you had on paper is actually what exists in the database.
