Every vendor in the support technology space is selling AI right now. Chatbots. Auto-summarisation. Sentiment analysis. Predictive routing. Agentic workflows. Co-pilot assistants. The pitch is always a version of the same thing: your team will do more with less, your customers will be happier, and the ROI will be obvious.
Some of that is true. Some of it is noise. And the ability to tell the difference — reliably, before you’ve signed a contract — is one of the most valuable skills a support leader can have in 2026.
I’ve been writing about AI in the contact centre since 2018, and I’ve spent the better part of the last four years implementing it in production. I’ve written about the agentic AI landscape and what it actually means for operational leaders. What I haven’t written is a straightforward framework for how to think about procurement decisions — what earns a place in your stack, what you can build yourself, and what you should leave alone entirely.
That’s what this is.
The question that matters before any other.
Before you evaluate a single tool, you need to be honest about whether your operation is ready to use it.
AI doesn’t fix broken processes. It accelerates them — which means a broken process with AI becomes a faster, harder-to-stop, broken process. If your ticket categorisation is inconsistent, AI triage will classify tickets at scale into an inconsistent taxonomy. If your agent skills aren’t defined and documented, AI routing has no signal to route against. If your training framework is informal and tribal, automated onboarding tools will automate the transmission of whatever your informal norms happen to be.
This is not a theoretical concern. The most important lesson from a large-scale AI triage implementation I’ve been close to was that the data preparation — auditing, cleaning, and standardising years of historical ticket data before the model was trained — took longer than the AI implementation itself. The AI was weeks. The data work was months. Skip that step and you’re training your model on noise.
The same principle applied to team structure. Before skills-based AI routing could work, the team needed to be rebuilt into defined tiers with documented competencies. The AI needed to know what intelligent distribution meant — and that required humans to define it first. The sequencing mattered as much as the technology.
So: before you evaluate what AI to buy, audit what your operation looks like without it. The readiness question is not “do we want AI?” It’s “is there a specific, well-defined problem here that AI is the right tool for?”
What to buy.
Buy AI tools when the problem has all of these characteristics: high volume, highly repetitive, well-defined output categories, and low tolerance for inconsistency.
Ticket classification and triage fits this exactly. A high-volume support operation processing hundreds of tickets daily can realise significant savings from a well-implemented triage model — but only if the categories are clean and the historical data is consistent. When those conditions exist, the economics are compelling. When they don’t, you’re buying a tool that will surface your data quality problem at scale.
Auto-summarisation is another strong buy for most mature operations. AI-generated ticket summaries, call transcripts, and case notes save agents meaningful time and reduce the variability in how context is captured. The ROI here is less dramatic than triage but more broadly applicable — it works even when your underlying processes aren’t perfectly defined, because it’s augmenting a human rather than replacing a judgment call.
Knowledge base deflection — chatbots or virtual agents that can resolve straightforward queries without human intervention — is worth buying when your knowledge base is well-maintained and your common query types are genuinely answerable without nuance. The caveat is significant: a chatbot is only as good as the knowledge it’s drawing from. If your knowledge base is out of date, fragmented, or poorly structured, the bot will confidently give wrong answers. Fix the knowledge base first. Then buy the bot.
What to build.
Build AI solutions when the problem is specific to your operation in a way that off-the-shelf tools can’t serve, or when you have the data and technical capability to create something meaningfully better than what the vendor market offers.
Custom routing logic is the most common example. The vendor tools for AI routing are improving rapidly, but they’re designed for general use cases. If your operation has unusual complexity — multiple brands, a tiered skills framework, non-standard escalation paths — you may find that the configurability of off-the-shelf tools hits a ceiling before it reaches your requirements. In those cases, building the routing logic on top of a more flexible platform (or in your CRM directly) will produce better outcomes than trying to fit a general tool to a specific problem.
Internal tooling for agents is another strong build case. AI-powered agent assist tools — suggesting responses, surfacing relevant knowledge base articles, flagging potential escalations in real time — can be highly effective when they’re tuned to your specific product, your specific customer language, and your specific resolution patterns. The vendor versions of these tools are general. A built version can be precise. If you have the data and the technical resource to build it, the precision is worth it.
What to skip.
Skip AI tools when the problem requires human judgment, empathy, or contextual reasoning that current AI genuinely can’t replicate.
Complex complaint handling is the clearest example. A customer who is angry, distressed, or escalating needs to feel heard by a human — and the emotional intelligence required to navigate that interaction is not something any current AI tool handles well. Deploying AI to manage escalated complaints is one of the fastest ways to turn a recoverable situation into a lost customer.
Skip tools that promise outcomes without being specific about mechanisms. If a vendor can’t explain precisely how their model works, what it was trained on, how accuracy is measured, and what happens when it’s wrong — that’s not a product you want in your production environment. Confidence in a demo is not the same as reliability in production.
Skip anything that requires your team to become significantly more sophisticated to operate. The best AI tools make the humans around them more effective. Tools that require heavy ongoing configuration, constant monitoring, and specialist knowledge to maintain are tools that will slowly degrade as your team’s attention moves elsewhere. Maintenance cost is a real part of the ROI calculation, and most vendors don’t put it in the slide.
A note on build vs buy sequencing.
The practical reality for most support leaders is that you’re not choosing between building and buying in isolation — you’re making those decisions in sequence, under budget pressure, with limited technical resource. A few principles that have served me well:
Buy to validate the use case before you invest in building it. If you’re not sure whether AI triage will work in your environment, a vendor tool with a short contract is a cheaper way to find out than a six-month build. Prove the value at small scale, then consider whether a custom build is warranted.
Clean your data before you do either. Whatever you buy or build, it will be trained on or evaluated against your historical data. That data is almost certainly messier than you think. The audit and cleanup is unglamorous, but it is the work that determines whether everything downstream succeeds or fails.
Involve your team in the evaluation. The agents who will use these tools day-to-day will see problems in a demo that management won’t. They’ll also be the ones whose trust you need to maintain through the change. AI implementations that are done to support teams rather than with them generate the predictable result: shadow workarounds, quiet non-adoption, and a vendor renewal conversation where you can’t point to much uptake.
The tools are getting better. The fundamentals of how to evaluate, implement, and manage them are not changing. Start with the problem. Define the success criteria before you look at solutions. Prepare your data and your team before you go live. And build in three to six months of active improvement after launch before you treat anything as complete.
That’s not pessimism. It’s just the actual work.
Hutch Morzaria is a Director-level CX and Support Leadership professional with 19 years of experience building global support organizations across SaaS, Fintech, and enterprise technology. He has hired dozens of support and CX leaders across his career and holds ITIL Expert certification across V3 and V4.
