What contact centre planners get wrong about gen-AI
The slide deck doesn’t match the operation
Every contact centre conference in 2026 has the same headline slide. Gen-AI will transform the operation. AHT will fall 20%. Deflection will hit 40%. Quality scoring will be automated. Forecasting will become real-time. The slide is being shown by vendors, consultancies, transformation teams, and increasingly the board. The planning team that has actually piloted these tools and measured the outcome knows the slide is wrong on most of its specifics, partly right on a few, and the gap between the marketing and the operational reality is wider than most operations are willing to acknowledge.
This article is the honest planner’s take. Where gen-AI genuinely helps, where it marginally helps, where it doesn’t deliver, what it does to AHT and deflection in practice, the forecasting and capacity-planning implications, and the leadership conversation that prevents the operation from over-committing to a transformation that won’t arrive in the shape promised.
Where gen-AI genuinely helps
Three application areas have produced consistent, measurable improvements across operations that have run honest pilots.
Advisor assist. Real-time prompts during a call — suggested next-best-action, knowledge surfacing, draft email responses for agents to edit. The well-implemented versions reduce AHT modestly (3–8%, not 20%) and improve quality scores by reducing the time agents spend hunting for information. The poor implementations distract advisors and produce a small AHT increase. The differentiator is the quality of the underlying knowledge base — AI doesn’t fix a bad knowledge base, it just exposes it more visibly.
AI-led QA scoring. 100% coverage of contacts instead of 4–6 evaluated by humans per agent per month. Most useful for compliance items, specific phrases, and surface-quality features (clarity, structure). Less useful for outcome and empathy. Operations that run AI-led QA alongside a smaller human QA programme (see AI-led vs human QA) typically catch more issues sooner and free up evaluator time for the contacts the AI flagged as interesting.
Knowledge surfacing. The agent asks a natural-language question and the AI returns the relevant policy excerpt with citation. Replaces the “hunting through a slow intranet” problem most agents live with. Saves 10–30 seconds per contact where it’s used. Modest impact on AHT, real impact on agent experience.
Where gen-AI marginally helps
Post-call summarisation. The AI drafts the wrap-up note; the agent reviews and accepts. Cuts ACW time by 30–60 seconds for the cases where it works; doesn’t work for complex contacts (which is where ACW time actually matters). Net AHT impact: small but real.
Forecasting model assistance. AI-augmented forecasting tools that pull in external signals (weather, marketing calendars, news events). Lifts model accuracy by 1–3pp WAPE on stable queues; doesn’t lift it meaningfully on already-good forecasting teams. The teams with the most to gain are the ones running unsophisticated forecasting in the first place.
Coaching content generation. AI-drafted coaching feedback based on QA scores. Useful for team leaders who are time-poor; risks formulaic feedback that lands badly with experienced agents.
Where gen-AI doesn’t deliver what the deck promises
The deflection myth. The slide says AI chatbots and voicebots will deflect 30–50% of contacts. The measured reality across UK contact centres in 2026 is closer to 5–15%, often with a negative side-effect: the contacts that do come through are harder, longer, more frustrated, and produce a meaningful AHT increase that offsets some of the deflection gain. The customer who’s already failed at self-service is in a worse mood than the customer who came straight to voice. Operations that scale deflection often find their voice queue’s total handle-time demand falls less than the contact-count drop suggests.
The full agent replacement story. No vendor pitches this directly, but the implicit story is that voicebots will gradually replace voice advisors. The measured experience is that voicebots handle a narrow band of simple, transactional contacts well and a broad band of complex contacts badly. Forty-year-old companies trying to replace experienced agents with bots are buying themselves a brand-damage incident, not a cost saving.
The “forecasting becomes real-time” pitch. Some vendors suggest AI will continuously re-forecast intra-day and adjust staffing in real time. The maths is correct; the operational reality is that you can’t move agents around in real time without breaking the schedule and frustrating the workforce. The signal isn’t useless — it can inform real-time decisions — but the “dynamic staffing” vision is decades from being operationally viable in any unionised or contracted workforce.
The end of the planning team. No vendor says this out loud, but it’s implied in the “AI does the forecasting” positioning. The reality is that gen-AI shifts what planners spend time on — less manual interval forecasting, more model curation, scenario modelling, and the interpretive work the AI can’t do. The planning team gets smaller in places and grows in others; the function doesn’t disappear.
Implications for forecasting and capacity planning
Three concrete shifts in how the planning team should think about gen-AI in 2026.
Plan capacity at conservative deflection. If the digital team is telling you 40% deflection, plan capacity at 10%. If the actual measured deflection comes in at 25%, you can adjust upward. Planning for 40% and seeing 10% leaves the operation with a queue you can’t cover.
Model the AHT effect, both directions. Advisor assist may cut AHT 3–8%; deflection may add AHT 10–20% on the remaining contacts. The net is operation-specific. Don’t take the vendor’s number for either; measure both yourself.
Treat the contact-mix change as the bigger planning problem. See capacity planning when the contact mix is changing. The compositional shift — simple contacts deflected, complex contacts remaining — is more important to plan for than the raw volume change.
What planners should actually do
Five practical moves that separate planning teams handling gen-AI well from those reacting to it.
1. Insist on a measurement framework before any pilot. The vendor’s number is not the number. Pre-pilot baseline, post-pilot measurement against the baseline, control segment if possible. Operations that pilot without a measurement plan never know what worked and what didn’t.
2. Be the voice of operational reality. The planning team is uniquely placed to know what an intra-day staffing change actually costs, what AHT really does to the queue, what 1pp of FCR is worth. When the digital team or the transformation lead is selling a number that doesn’t pass the operational smell test, the planning team is the one with the credibility to push back.
3. Pilot the high-value applications first. Advisor assist, AI-led QA, knowledge surfacing — the three areas that consistently produce real benefit. Defer the more speculative applications (voicebot deflection, automated forecasting) until the foundational ones are working.
4. Train the team to use AI tools without becoming dependent on them. The forecaster who uses AI-augmented tools should still understand the underlying maths; the analyst who uses AI-generated coaching should still recognise good and bad coaching. The skill remains; AI is a multiplier on it.
5. Track the cost honestly. Gen-AI tools cost real money — both the platform subscription and the integration effort. The ROI conversation has to include the cost, not just the savings. Operations that count only the benefit end up with vendor spend that quietly accumulates.
The leadership conversation
The transformation team or the operations director will, at some point, ask the planning team to commit to a capacity reduction based on AI-driven savings. The honest answer needs to be ready:
“We’ll re-plan capacity based on measured outcomes, not committed benefits. The pilots we’ve seen show real but smaller improvements than the vendor pitches. The risk to the operation of over-committing is much larger than the cost of being slightly cautious. We’ll catch up to any genuine benefit within a quarter once it’s measured; we can’t catch up from an SL miss caused by under-staffing against benefits that never materialised.”
That answer is unpopular with the transformation team and right for the operation. The planning team that says it builds credibility with operations and CFOs over time; the planning team that doesn’t spends the following year explaining why the savings didn’t arrive.
The realistic 2-3 year picture
By 2028, the honest expectation for most UK contact centres is: advisor assist deployed and producing 3–8% AHT savings; AI-led QA running at 80% coverage with humans on the contested 20%; knowledge surfacing as a standard agent tool; voicebot deflection at 10–20% on simple-transactional flows; forecasting models augmented by AI but still curated by planners. That’s a meaningful operating-model shift; it’s not the transformation the deck promises.
The operations that get this right will be the ones that piloted carefully, measured honestly, scaled the proven applications, and resisted the pressure to commit to vendor numbers. The operations that don’t will spend the next few years explaining why their AI strategy didn’t pay back.
Conclusion
Gen-AI is real, useful, and over-sold. The planner’s job is to separate the signal from the marketing, plan capacity against measured rather than committed benefits, pilot the three or four applications that genuinely work, and have the unpopular conversation with leadership when the savings being claimed don’t match the operational reality. Done well, gen-AI is a steady accumulation of small operational improvements. Done badly, it’s a commitment the operation can’t deliver and a brand-damage event waiting to happen. The planning team that handles it well is the one quietly making the difference.
Pair this with AI-led vs human QA, AI for forecasting, capacity planning when the mix is changing, forecast with ranges, and composite metrics that hide the truth.