Collections is a dialogue, not a script

Most AI companies selling into finance are chasing the same promise right now: full autonomy. The agent reads the invoice, drafts the chase, sends it, and books the payment with no person involved. It demos beautifully, and it is the first thing a prospect asks to see.

We build the opposite by default. Spinal's collections agent gathers the context and drafts the action, then it stops and waits for a person to hit send. Full autonomy is available, one setting away, but customers have to ask for it rather than receive it out of the box. Here is why we think the less impressive demo is the right call, and why the customers who have used both tend to agree.

It isn't really about trusting the model

The common assumption is that finance teams resist AI because they doubt it will get the numbers right. That is not what we hear when we actually sit down with them.

The real objection is about being blindsided. A finance director in construction described the situation that worries him. A client rings up annoyed about a chaser, and no one on his team even knows which message the client means, because nobody on their side ever saw it go out. And telling the client "the AI sent it" is not an answer a finance team can give.

That has nothing to do with model quality. It is about accountability. When an autonomous agent sends something a customer reacts badly to, the people responsible for that relationship cannot reconstruct what happened or stand behind it. The message went out under the company's name, into a live commercial relationship worth far more than the single invoice, and the team that owns the relationship never saw it. A more accurate model does not close that gap. A person reviewing the draft before it sends does. The point of the human is not to catch arithmetic errors. It is to keep a name attached to every message that leaves the building.

This is also why a clean audit trail matters as much as the review step itself. Even with a person approving each send, finance teams want to look back and see exactly what went out, to whom, and on what basis. The human in the loop and the record of their decisions are the same trust mechanism viewed from two angles.

Chasing is a dialogue, not a script

This is the part most automation gets wrong, and it is the real reason we hold the line on human review.

Collections done badly is a fixed ladder of reminders. Day 7, a polite nudge. Day 14, something firmer. Day 30, a final notice. The software neither knows nor cares why the invoice is late. It climbs to the next rung and sends the next template, regardless of what has actually happened on the account.

Most late invoices are not refusals to pay. A large share of them are ordinary friction. A purchase order number that does not match the invoice. A line item the customer is still querying. An approver who is on leave for two weeks. A buyer who is themselves waiting on funds from further up a contract chain and genuinely cannot pay yet, no matter how firmly you chase. Each of these calls for a different response, and some of them call for no chase at all. The finance director we spoke to was blunt about what happens when a reminder ignores that history. The customer switches it off. In his words, a good chaser has to be a proper dialogue, because there are always two sides to a story.

That phrase is the whole game in receivables. The person on the other end usually has a reason, and often a good one. A chase that lands as though the previous twenty emails never happened does more than fail to collect. It signals to the customer that nobody on your side is actually reading, which is the quickest way to burn the goodwill that gets you paid this quarter and the one after. In sectors where the same client sends you work for years, that goodwill is worth more than any single overdue balance.

Keeping a person in the loop is what lets a chase carry that context. The agent reads the full thread, pulls the promise to pay from three weeks ago, notices the open dispute, sees that the buyer is stuck in a pay-when-paid chain, and drafts a message that reflects all of it. A person then checks that the draft matches the relationship before it goes out. The history lives in the data, and the agent is good at surfacing it, but the judgment about whether the tone and timing are right still belongs to someone who owns the account. Take the person out and you do not get a faster version of that judgment. You get a confident message that treats every late invoice exactly the same way.

Trust is earned per customer, not assumed

None of this means a person stays in the loop forever. It means autonomy is something a customer arrives at rather than starts with. They watch the agent draft chases under approval for a few weeks. They see the kind of message it writes and the context it picks up on. At some point they decide they would rather not review every routine reminder, and they ask us to let the agent send a defined set of them on its own. That decision belongs to the customer, and it tends to land in a different place for each one.

What this looks like in Spinal

The agent does the work and you keep the judgment. An issues tab surfaces the items that need a decision. An activity log records everything the agent has done on your behalf, so there is always a clear answer to what happened and why. Autonomy gets switched on later, once a customer has watched the agent work and decided for themselves that they trust it. We treat that as something earned with each customer rather than assumed on day one.

‍