A desktop sales AI tests when to decide—and when to hand over

Keep the arithmetic outside the model

Business-to-business sales quotes can change repeatedly: customers alter quantities, request longer payment terms or negotiate delivery dates. Each change can affect profitability and whether a salesperson has authority to approve the deal.

In an arXiv preprint, Yu Liu and Xingbang Du present LoMASA, a sales assistant running on one desktop computer. Their simulation study asks whether language models can take over some quoting decisions without making commitments riskier than a straightforward pricing-rule baseline. It is a software experiment, not a field trial with working sales teams.

The system separates calculations from judgment. Code computes prices, financing costs, margins, capacity and alternative offers. Its example policy targets a 15% margin, imposes a 5% minimum and limits salesperson-authorized discounts to 8%. A quote below the current margin floor, or above the discount limit, must go to a manager regardless of the model’s answer.

That protection concerns the current calculated state. It cannot guarantee that an initially acceptable quote will remain profitable after costs change.

Confidence determines the handoff

A 35-billion-parameter mixture-of-experts model handles the first decision stage. Such models activate only part of their parameters for each input; this one uses about three billion active parameters. It chooses whether to commit, commit conditionally or hold, and separately whether to escalate.

Rather than asking the model to describe its confidence, LoMASA reads the probabilities assigned to the available answer labels. Uncertain questions pass to a second, 27-billion-parameter model. If confidence remains insufficient, the salesperson decides.

Eight similar historical cases, retrieved using numerical features such as margin and recent cost changes, provide additional context. Model weights remain unchanged, although confidence adjustments and handoff thresholds are selected using validation data. Here, “training-free” does not mean the system needs no development examples or tuning.

The second model also drafts replies and manager briefs. A numerical validator checks that numbers in the text match permitted values, allowing rounding and small counts. A further model check looks for promises exceeding the decision. These safeguards are narrower than a guarantee that every sentence is correct.

What the simulated test found

The researchers assembled episodes from public supply-chain and retail datasets. Tomorrow’s cost supplied the hindsight needed to judge whether a commitment would remain acceptable. Training, validation and test splits used separate supply-chain games and customers.

The test contained 500 episodes and 1,512 decisions. Hard rules routed 42.5% directly to a manager; 19.5% required salesperson decisions. The remaining 38.0% were reported as automatically settled, with a median first-stage decision latency of 2.67 seconds on a Mac mini M4 Pro with 64 GB of memory.

“Automatically settled” needs qualification: the paper’s comparison of those 574 decisions includes 53 model-requested manager escalations. The category therefore should not be read as entirely human-free resolution. Nor is the first-stage latency a measurement of a complete customer interaction.

On those same 574 cases, LoMASA and the rule baseline both achieved 89.4% commitment-decision accuracy. Commitments whose margin subsequently fell below the floor numbered 15 for LoMASA versus 26 for the baseline—reported as 2.61% versus 4.53% of the comparison cases.

The reduction came through escalation, not better overall commitment accuracy. Across cases without mandatory escalation, the system generated 90 unnecessary manager referrals. Of 49 risky cases outside the hard rules, it escalated 11, deferred 23 to the salesperson and missed 15.

Useful confidence, limited second opinions

The first model’s answer probabilities tracked correctness better than confidence it stated in words. That matters because selective automation depends on identifying which decisions to leave to people, not simply maximizing average accuracy.

The slower model contributed little to decision-making: it independently settled only 1.5% of commitment questions across the full test. The authors suggest its review stage could be skipped, retaining its writing role. They did not test whether the first model could write equally well.

Important limits remain. Decision tests used the true structured events rather than parsed messages, so the main results do not establish end-to-end reliability. Perfect parsing on 600 templated, single-field messages does not demonstrate performance on messy customer correspondence. Only 12 test decisions involved cost jumps.

Human handling time and manager workload were not measured, so proposed hour savings are calculations based on assumed handling times. Testing used one hardware setup and one deterministic run, and the preprint leaves its code-repository link unfinished.

The result is evidence for a bounded design principle: let code enforce explicit constraints, use confidence to allocate judgment, and count the work transferred to people—not just the work removed from a salesperson’s queue.