Part I of our 'Scaling AI at Allica' series set out the three levels of our AI strategy: organisation-wide adoption, an AI-first approach to product and engineering, and building complex agents. Part II dived into the product engineering aspects, and IIa then shared detailed data tracking progress on those.
Part III is about agents - in this blog we will cover our growth team, and in a subsequent blog we’ll cover lending.
Why our growth team is critical to Allica’s success
'Go-to-market' with the established business segment we target is hard. By definition these customers have been successfully operating for a number of years and so have existing banking and financial providers. They may not be super happy with these but rarely are they actively looking for something new, as they are busy running their businesses.
Allica also offers a broad range of products for our target segment. A customer may first come to us looking for somewhere to hold cash, then want to take advantage of our expense cards, and then want to borrow to acquire another business. We interact with customers extensively through digital channels, through human customer success teams and relationship managers.
Making established businesses aware of Allica, converting them into customers and deepening those relationships over time is therefore a multi-channel, multi-touch approach, rather than a simple digital funnel.
Our growth team spans the whole lifecycle - acquire, convert, onboard, activate, serve and deepen - across channels, and across business accounts, cards and all types of lending. We have a dedicated centralised growth team, organised into roles that own different parts of the journey and can move quickly inside it. Over the past year this team has developed and put into use more than 25 agents and skills focused on different customer lifecycle stages.
The combination of breadth of lifecycle coverage and specialisation within it has worked for us, but has also created a problem.
More agents did not automatically make us smarter
One team owns acquisition, another conversion, another onboarding, another activation and another the ongoing relationship. There are also product splits. This has served us well: decisions are faster, measures are clearer, and nobody waits for a central committee.
But customer behaviour goes beyond that structure. An objection raised during onboarding may reveal that an acquisition message is attracting the wrong intent. An in-life campaign may find language that works better than the original landing page. A servicing conversation may expose a need that should change how a relationship manager prioritises tomorrow.
Those learnings matter across teams and agents, and how these are incorporated needs to be systematised. Traditional approaches such as sharing slides/documents in a regular team meetings is at best patchy, unreliable, and too infrequent to make learnings reusable. Before another team can act, a learning must be understood in its original context, tested for relevance, and translated into a different customer moment.
At the pace we were operating, that translation happened inconsistently. The volume of context had simply grown beyond what people could reliably connect in their heads - and every new agent made it larger.
Introducing: AGIS
So the team built AGIS: the Allica Growth Intelligence System.
This is both the collection of all the production growth agents we have built, and the connective layer that lets their context, actions and outcomes loop to improve one another. It was built to solve three specific problems.
-
Compounding what we learn. AGIS contains what the growth team already knows: the underlying need, pain points and jobs to be done; the objections and behaviours we have observed; the experiments already run; and what happened afterwards.
It interprets that evidence using organisational, product and go-to-market context, as well as the specific team's own context and processes. A team beginning something new therefore starts with the relevant learning, its source and level of confidence, rather than another team's document or a blank page. This is the most developed part of AGIS today.
-
Deciding where effort should go. Once a team has many agents and experiments running, activity becomes a poor proxy for value. An agent can generate hundreds of outputs without moving anything.
AGIS therefore has a live, governed outcome driver tree. Every skill, agent, insight, observation and experiment attaches to the metric or metrics it is expected to affect, and the tree links those measures to the relevant observed commercial outcome. That gives us a consistent view across experiments: which capabilities were used, how often, which metrics moved and what happened to the downstream outcome.
Too many things move at once to give us perfect causal attribution, but it lets us identify strong relationships, compare effort and decide where to invest next.
-
Suggesting what to try next. As AGIS accumulates context across the lifecycle, it can propose connections no single team is likely to hold at once. For example, a recurring objection that maps to a particular acquisition segment, or an in-life message that suggests a better way to frame the proposition much earlier.
The decision to adopt a change is still human-controlled given it drives changes across the journey, and we want to ensure the optimisation loop is not just converging on what it already knows how to improve (i.e. a local maxima optimisation issue).
The range of what sits inside AGIS is broad. At the top of the funnel, agents turn fragmented market and company information into a usable view of each prospect's fit, potential value and likely needs. In acquisition and conversion, scoring and prioritisation help colleagues decide who needs attention now.
Human-customer conversations are analysed to provide comprehensive and immediate feedback for coaching. For campaigns, agents underpin research, drafting, variation and quality checking. After a customer onboards, timing signals identify where a human intervention might aid activation, prevent avoidable attrition, and the next best action. Underneath all of it, reporting and validation agents assemble evidence, reconcile data, and flag where a result is strong, weak or too early to read.
Some of these are sophisticated agents, some are narrow skills inside a workflow; some combine deterministic automation with AI judgement. We care less about the label than about whether the capability makes a better decision possible, and whether we can check afterwards what it did.
The enabling layer is crucial: AGIS is managed as a versioned repository where context and experiment records have owners and sources, changes are reviewed, reporting is validated before it becomes shared context, and access and redaction are handled properly because some of the underlying evidence includes customer conversations and colleague-level information.
The shared memory needs maintaining to be optimally useful. This mirrors what we learned in our product and engineering teams: unlocking the combination of velocity and quality improvements needs the right harness around it.
One learning, several teams
The clearest example of the compounding loop started with a hypothesis we nearly acted on.
Applications were stalling, and a rate-boost feature on the account looked like the reason. It came up frequently in customer calls. The obvious fix was to explain it better, and we were close to briefing that work.
Instead we looked at more than 1,200 lost or stalled journeys, with usable conversation data for around 87% of them. Fewer than 3% turned out to be genuinely blocked by how the feature worked. Two much larger issues sat underneath. The first was a mismatch between the intent a customer arrived with and how we were presenting the wider proposition including the boosts - people had come looking for one thing and been offered something broader. The second was a separate group we had never had a meaningful conversation with at all.
One apparent problem was two different problems, with a third much smaller one attached. Had we acted on the original hypothesis, we would have fixed the smallest and left the other two roughly as they were.
They went to different teams and produced different work.
The proposition finding changed how we talk to customers arriving with a narrower need: meet the intent they came with, then bridge into the wider proposition, rather than leading with everything at once. A re-engagement campaign built on that principle saw its best-performing message drive roughly half of all engagement and returned several opportunities to active pipeline.
The contactability finding went down a different path - into routing, coverage and giving customers the option to self-serve book a convenient time rather than waiting for our customer success team’s next attempt to reach them. In the first full month, the share of qualified opportunities marked unreachable fell 16%, and bookings rose sharply against a prior baseline.
This might be described as a self-learning system. We use the phrase carefully. AGIS does not rewrite its own rules or independently launch changes that affect customers. Learnings are captured, sourced and reviewed, and people decide whether the evidence is strong enough to change the next action. The system compounds because it reduces how often anyone starts from zero - not because judgement has been removed.
Examples of agents
Within that shared intelligence layer, various growth agents are running to deliver speed and feedback. Examples include:
From audience to output. Producing a campaign used to mean briefing a list, waiting for it, writing copy, and reviewing it. Now a marketer can ask 'ABM Brain' for an audience such as cash-rich prospects in a target sector, whether or not they have interacted with Allica before, and the targeting agent returns the relevant group with the context behind it.
A drafting agent writes from the product marketing playbook, so a draft of the copy arrives already in Allica's tone of voice and inside brand guidelines rather than being corrected into them. A review agent checks the draft and generates variations, turning one approved message into a testable set rather than a single static asset.
The same pattern works after onboarding: a journey skill turns an agreed customer objective into the right cadence and triggers, and performance feeds back into the next audience or cohort, message and treatment. The team still decides what is worth saying and what reaches a customer; the agents remove much of the gap between deciding and testing.
The coaching loop. Conversation coaching used to run monthly, from whatever sample of calls a manager had time to listen to. Agents now read the whole archive. In a single month we analyse over 2,000 conversations, giving valuable feedback for the customer success executive or central relationship manager. Propensity models help colleagues prioritise proactive calls to prospects based on fit, engagement and opportunity, as well as to existing customers with an identified need.
A simple example came from onboarding conversations. Colleagues proactively asked prospects how they had first heard about Allica in fewer than half of the relevant calls. As a result, around a quarter lacked enough information to identify indirect influences or correct the original attribution. This is a coachable behaviour and an important part of collecting the data needed to evaluate and optimise multi-channel acquisition.
What this adds up to
AGIS is not finished, and we would not attribute every movement in our growth team's performance to it. Agents operate alongside great people, product development, and processes. In aggregate we see powerful results though:
- Across acquisition, a redesigned partner journey increased its form completion rate by over 40%
- At portfolio level, average monthly account openings grew 18% vs the benchmark period
- Weekly prompts helped relationship managers prioritise newly onboarded customers for outreach. Contacted customers were over 25 percentage points more likely to start making payments, although other prompts showed little or no measurable difference.
While we seek to have clear attribution via the driver tree nodes, this ultimately cannot be a perfect science - in most growth team work, isolating a single cause is hard: a prioritisation agent operates alongside new training, different routing, revised messaging and a product change. Where it is practical we run controlled tests; elsewhere we look for repeated movement in the specific driver an intervention was built to change, compare the most relevant groups we can, and state the likely selection effects.
Where we go next
Compounding learning is working. Driver-tree measurement and prioritisation are live and being optimised; they let us compare agents, skills and experiments consistently, even though we do not expect perfect causal attribution. Some agents work well, some will be replaced as models improve, and some experiments have produced little measurable benefit.
The destination is increasing autonomy across the growth team's loop: an outcome changing somewhere in the lifecycle should surface the relevant customer context and prior experiments, propose the next test, help build it, and put it in front of the person who holds the authority to run it.
But any optimisation loop converges on what it already knows how to improve. A system compounding its own learning will, left alone, get better and better at the growth model it already has - and stay there. People are needed not only as a safety control but to challenge the objective, introduce a hypothesis the evidence does not yet support, and work out when the next source of growth sits outside the map the system has built.
That is consistent with the principle we set out in Part I. The value of agentic AI comes from letting it run more of a complex process, with the data, workflow integration and governance to make the result useful. The judgement about where Allica should go, and what is right for our customers, stays with people. We help optimise human performance, rather than replace it.
