Most organisations deploy AI to deflect contacts – and then stop. That skims the smallest of three layers of value. The assist and resolution layers are harder, less visible and worth several times as much – and the order in which you build them decides whether your programme compounds or plateaus.

When a leadership team says “we’re using AI in customer service”, it almost always means a deflection bot – a chatbot or an IVR that handles simple enquiries so they never reach an agent. Deflection creates real value, but it is also the shallowest of three layers on which AI compounds in customer service. Teams that stop at deflection capture perhaps a fifth of what is available, and then wonder why their AI programme has stalled.

We have built on all three layers for contact centres and operations teams, and the value grows by roughly an order of magnitude the higher you go – as does the difficulty. Which is why most stop at the bottom.

The real problem: deflection gets measured, the bigger layers do not

Deflection has an obvious, satisfying metric: contacts deflected × cost per contact = money saved. It is easy to measure, easy to put on a slide – and capped: you can only deflect so many contacts before you start deflecting ones that needed a human. That generates callbacks and erodes trust. The assist and resolution layers are worth more but are harder to measure, and are therefore underfunded. Investment follows measurability, and what gets measured is what is easy to count.

The three layers

Layer 1 – deflection (catch the contact)

The AI fully handles routine, self-service-style enquiries: balance, status, simple changes, FAQ answers. Value: avoided contact costs. Ceiling: low – and pushing past it achieves the opposite. Over-deflection sends people who needed help into a loop, and they call back angrier. Think of deflection as “automating genuine self-service”, not as “blocking people”.

Layer 2 – assist (make every agent faster and better)

On every contact it did not deflect, the AI works alongside the human: real-time response suggestions, pre-drafted replies, automatic conversation summaries, surfacing the relevant policy or an earlier case. This touches the contacts that require a human – the majority – and improves handling time, consistency and quality across all of them. It also raises the floor: your newest agent answers more like your best one. The value here is typically a multiple of deflection, because far more contacts are touched and because it lifts not just cost but quality.

Layer 3 – resolution intelligence (fix why contacts happen)

The AI reads the entire contact stream to work out why people get in touch, and feeds those insights upstream to remove the cause. If 8% of contacts concern a single confusing line on the bill, the most valuable action is not a better bot for that question – it is a corrected bill, so that those contacts never arise. This layer attacks contact volume at source and improves the product, not just its support. The value here is the largest and the most strategic, because reducing demand beats handling demand.

A worked example: the three-layer build at an energy utility

An energy utility with around 120,000 contacts a month started, as usual, with a deflection bot.

• Deflection: caught around 18% of contacts (meter readings, simple account queries). Saving ≈ CHF 1.1 million a year. Real – and the point at which most would declare victory.
• Assist: draft replies, summaries and answer surfacing for agents on the remaining 82%. Average handling time −19%, after-call work −40%, and the spread in quality scores between new and experienced agents narrowed markedly. Value ≈ CHF 3.4 million a year from capacity gains and shorter onboarding – roughly three times the deflection layer, and on contacts deflection could never have reached.
• Resolution intelligence: analysing the contact stream showed that around 11% of all contacts traced back to two fixable causes – an unclear section of the bill and a broken self-service password reset process. Fixing those at source eliminated contacts permanently – estimated value CHF 2.6 million a year and rising, plus a measurable improvement in customer satisfaction, because the best contact is the one the customer never had to make.

Total value roughly six times the pure deflection programme that was originally planned – and the resolution layer kept paying out after the project ended, because it removed demand instead of handling it.

The lesson about sequence: deflection created credibility, assist delivered the bulk of the recurring value, and resolution intelligence turned a reactive operation into a preventive one. Each layer reused the data and integrations the previous one had created – so the build compounded instead of starting over each time.

Why the layers have to be built in this order – and on a shared data backbone

The three layers are not independent products you pick à la carte; they are a sequence that reinforces itself – and that effect only works if you build the foundation once. Deflection forces you to capture your most important intents and connect your systems of record. Assist reuses exactly those intent models and integrations, now pointed at the agent instead of the customer. Resolution intelligence uses the contact stream and the intent taxonomy that the first two layers have already produced. Build them on separate stacks and you pay the integration tax three times and get no reinforcement at all; build them on a shared data backbone and each layer costs less than the one before.

This is also why jumping straight to the resolution layer rarely works, even though it is the most valuable. Without the instrumentation that deflection and assist bring – structured intents, clean transcripts, consistent tagging – the contact stream is too noisy to analyse reliably. The sequence is not bureaucratic caution; each layer produces the data the next one needs. The mistake is to treat them as alternatives (“should we do a bot or analytics?”) rather than as a staircase where every step shortens the next.

One organisational warning decides whether the resolution layer ever becomes reality: it is the only layer that pays off outside the contact centre. Correcting the confusing bill is finance’s job; fixing the broken password process is the product team’s job. If the AI programme belongs to support alone, the resolution insights have no recipient and die quietly in a dashboard nobody acts on. The organisations that get value from this layer give it an executive sponsor who can task other areas with the fixes – turning “this is why customers contact us” into changes that actually remove the demand. Without that mandate, you generate the most valuable insight of the whole programme and watch it gather dust.

Pitfalls

• Stopping at deflection. The easy metric is the small prize. Plan for assist and resolution from the outset.
• Over-deflection. Catching contacts that needed a human creates callbacks and annoyance; measure resolution of the enquiry, not just containment.
• Assist that slows agents down. If the suggestions are noisy or wrong, agents will ignore them. Optimise for precision and let agents dismiss suggestions easily.
• Ignoring the resolution layer because it is cross-functional. Fixing root causes takes product and operations, not just the contact centre – which is exactly why it gets neglected, and exactly why it is so valuable.
• No shared data backbone. Building every layer on its own infrastructure throws away the reinforcement effect; design the data foundation once.

Decision checklist

1. Are we measuring resolution rate and callbacks, or only deflection and containment?
2. Have we planned the assist layer for the contacts that require a human?
3. Are we analysing the contact stream to find root causes and fix them upstream?
4. Does assist improve quality and consistency, not just speed?
5. Is there a shared data foundation, so that each layer builds on the previous one?
6. Does ownership of the resolution layer sit cross-functionally with product and operations, not just with support?
7. Do we know the relative value of each layer for our own contact mix?

The gridz view

Deflection is where AI in customer service begins – and the worst place to stop. The compounding value lies in making every human contact better (assist) and in removing the reasons contacts happen at all (resolution intelligence). These layers are under-built precisely because they are harder to measure and cross more organisational boundaries. We build all three on a shared data backbone, in a sequence that funds itself, so that the programme keeps paying out instead of stalling at the first easy metric.

If your AI in customer service is a deflection bot and nothing more, you have built the foundation and stopped on the cheapest floor of the building. The next step is not a better bot – it is AI for the agents who handle everything the bot could not, and after that the analysis of the contact stream for the reasons those contacts exist in the first place. The multiples sit in those two steps, and they use the data and integrations you have already paid for. The best contact your customer ever has is the one she never had to make – and only the resolution layer gets you there. If you want to plan that next step, get in touch.