Deciding to cut an underperforming AI tool is the easy part. Getting out of the contract can be painful for CIOs. Vendors made onboarding easy with generous trials, quick setup and AI included at every level. However, these same vendors often make leaving costly and slow. Several contract provisions can complicate or increase the cost of an exit: 

When you factor in the challenges of untangling integrations, moving data and retraining teams, what seems like a budget win can turn into a lengthy and expensive process. 

Before you cancel a single license, it's important to know where the pitfalls lie and how to navigate them. Here's what to check, what to contest and how to leave without paying more to exit than you did to stay.

Shadow AI stumbling blocks

Do you know where the vendor has been inside your operation, and what might break if you rip it out? If you don't, then find those details first.

Related:Before the next VMware: How CIOs prepare for vendor shocks

"Map every internal process that touches the vendor before you give notice. This sounds obvious, but often isn't done. Shadow usage, employees who adopted a tool on their own because the official tooling was slow, is where disruption usually comes from, not the formally sanctioned workflows," said Frank Meltke, CEO of Contraco, a global digital transformation consultancy.

Another area you should evaluate carefully before initiating an exit is the why behind every shadow AI tool's selection. If the answer is consistently "because the approved option was too slow, too restricted or didn't exist," the exit strategy should "include fixing that, or the same AI tools will come back through a different door six months later," Meltke said.

First steps toward the exit

Start at the beginning, so you know where you stand now. That is, start with a careful read of the vendor contract, paying special attention to the data retention, training rights and exit terms. 

Next, map workflow dependencies before decommissioning. "Most exits fail because the dependency map wasn't built first. Sequence the exit so critical workflows have replacement paths before vendor access is cut," said Diptamay Sanyal, principal engineer at CrowdStrike.

 Paul McDonagh-Smith, visiting senior lecturer in IT at the MIT Sloan School of Management, recommends treating vendor exits as staged migrations rather than one-time cutovers.

"Proceed like a surgeon, not a butcher: map dependencies, shadow-run replacements for one cycle before decommissioning and negotiate from the renewal date backwards. Sequence by reversibility, working on straightforward exits first, the load-bearing ones next," advised McDonough-Smith, who is also senior advisor to NASA Goddard Space Flight Center. 

Related:Priceline CTO prioritizes engineers able to 'hold a room and a roadmap'

Putting AI vendors on notice

Take care how you handle that vendor notice, too. 

"The lowest-disruption path is to align vendor exits with renewal cliffs rather than midstream terminations," said Jackie Swanson, managing partner at Gartner Consulting. 

Just don't let automatic renewals get past you. You'll also want to begin your exit well ahead of the targeted end date. 

"Start the exit process 90 days before you intend to terminate, not 30," Meltke said. Most enterprise AI contracts have automatic renewal clauses, data retention windows and API deprecation timelines that "require lead time to navigate cleanly," she explained.

However, you don't have to wait until the end of a vendor contract to call it quits. Some contracts permit termination for convenience, usually for the unused subscription fees, according to John Pavolotsky, partner at Stoel Rives LLP and co-chair of the law firm's AI, privacy and cybersecurity group. "Even if a contract does not permit termination for convenience, nothing precludes a tool owner to call the vendor and work out an amicable exit," Pavolotsky said.

Related:IT leaders should never let a good crisis go to waste

Negotiating offboarding terms

As for offboarding terms, they should have been settled and documented at the beginning of the vendor relationship. Data deletion timelines, model fine-tuning ownership, output log retention and other pertinent details should all be spelled out in the initial contract. 

"If they weren't, you're in a harder position, but you can still request written confirmation of deletion with a specific date and a compliance statement from the vendor's legal team. Get it in writing before the final payment clears," Meltke said.

Ditto for clauses in the contract pertaining to data ownership and audit protections. If they aren't spelled out from the beginning, you may have unresolvable problems now. Even so, it's worth the effort to try to secure those now.

"Ask specifically about prompt and completion logs, any fine-tuned or adapted model weights derived from your data, cached API responses and any data passed to subprocessors or third-party infrastructure providers," Meltke said, adding that the last category is where "most gaps live." 

Be aware that the primary vendor may delete your data while a monitoring or analytics subprocessor retains it. "If you're in a regulated industry, vendor confirmation should reference whatever compliance standard applies -- GDPR Article 17, HIPAA or sector-specific frameworks -- so there's a clear audit trail if you're ever asked to demonstrate compliance," Meltke said.

Be very diligent in protecting every bit of your data because any and all of it is coveted. AI vendors are "absolutely ravenous for data, and if they can keep your data, they will," said Ranjith Raghunath, CEO of CX Data Labs.

Also beware of vague vendor promises that are not explicit. 

"We'll delete it is a claim to verify, not a promise to accept," McDonagh-Smith said. He added that it's important to settle three points in the contract before signing the initial contract or the last check before you exit a contract: 

Putting exit lessons to work in new AI contracts

Stay professional and keep notes while exiting AI vendor contracts, as you may need to re-enter agreements with them later. If you do reinitiate contractual agreements with them at some point, be sure to keep your options open.

"The vendor landscape is in the middle of a consolidation cycle. Contracts signed today will look very different in 18 months, and the smart CIOs are building optionality into every new commitment," Swanson said.

Keep in mind that strategically culling AI is not a rejection of AI but rather a cost- and risk-control measure typically applied to any technology investment. This also protects capital that will be needed later as AI improves and the business itself evolves. 

So when is the right time to reinvest in more or better AI? 

"Reinvest when discipline returns, not when the budget rebounds," McDonagh-Smith said. In other words, "scale up when the cause of the sprawl is resolved, not when the budget recovers," he said.

Three conditions must exist first, according to McDonagh-Smith. 

"Then mind where the money goes: the pull is always toward safe efficiency plays, but advantage lies in integration and innovation. Let demonstrated demand drive investment, not the fear of missing out," McDonagh-Smith said.

Exit terms will become a much larger issue over time. That's especially true if your company uses AI to create its own tools or commits its own processes and expertise to AI vendor platforms. Your company's workflows and agents can easily become dependent on the platform where they are stored. 

"If you cannot move your agents and your workflows from one platform to another, you risk a very significant loss, similar to the loss of an entire team, when you move to another platform," said U.K.-based Richard Nicholas, AI partner at law firm Browne Jacobson.

"This gives your current provider little incentive to keep costs down! It is well worth getting this agreed as part of the deal," Nicholas said.