Frontier Intelligence for the Frontline Sales Force

A $2B+ revenue Indian consumer goods business whose products reach 290 million households came to Bain & Co and Hyde to build an AI-powered frontier intelligence system. General trade, like at a neighbourhood kirana store, accounts for more than half of company revenue. Within it, CPG Co. had direct access to nearly 2 million retail outlets, served by a pan-India field force of 5,000+ distributor sales representatives (DSRs) who work every day to position and sell a portfolio of 500+ SKUs across eight categories in packaged foods and beverages.

A significant portion of their business is negotiated at a shop counter, in conversation between DSRs and retail store owners. Every one of those conversations carries signal: which pack sizes sell quickly, what margin a competitor is offering this fortnight, and why a newly launched product has not been restocked. This is the richest market intelligence the company generates, and none of it was being captured: the conversation happened, the DSR moved on, and the signal was lost. Hyde and Bain solved this problem together, operationalizing intelligence from these daily conversations by turning it into a compounding asset that could deliver significant topline growth for CPG Co. In partnership we built and delivered Saathi (meaning 'companion') for CPG Co. — the AI field intelligence tool that closes the loop between the conversation at the shop counter and the company's strategy.

What we built

Saathi has three key components. Wearables capture the retailer conversation. Hyde's speech and reasoning models process thousands of transcripts a day and produce a sales strategy for each outlet. Finally, a voice agent shares the recommendation with the DSR in their own language, in time for their next store visit. All of this works within CPG Co.'s existing software interfaces, and the workflow scales to the whole organization.

Turning brand ambition into daily instructions

The marketing team assembles monthly sales strategies from multiple data sources: primary-versus-secondary sales as a proxy for channel stock, third-party retail audit panels, and competitive intelligence fragments sourced from DSR managers. These broad sales strategies are turned into daily instructions: must-sell product lists and active retail promotions. DSRs access these recommendations through their Saathi app, which also tracks their activity, creating a two-way communication channel between managers and sales reps.

Where frontier models fail

Recording the conversation is easy. Recovering the signal inside it is not: the audio is accented, switching between local languages and English, captured at conversational pace in a noisy shop, and dense with SKU names and trade shorthand that frontier speech models do not recognize. Answering unscripted follow-ups in a conversation with a DSR requires our speech agent to assemble an answer utilizing current information about the rep, the store they're in, and the entire product catalog — fast enough that the conversation doesn't stall. All of this has to run at scale for 5,000+ DSRs tracking nearly 2 million retail stores (amounting to roughly 50 million minutes of voice a year), at a cost structure that works for the CPG company.

How we won

Solving this problem required us to create a scalable pipeline that transcribes raw audio and parses the transcript to prepare spoken recommendations, leveraging CPG Co.'s existing data.

Speech recognition built for the real world of kirana retail

Off-the-shelf speech-to-text degrades sharply on this input, so we trained custom STT models on in-domain audio: rep and retailer exchanges in local languages, code-switched, at pace, under shop-floor noise. The focus was on the vocabulary that matters downstream; for example, the model needs to transcribe ashwagandha correctly and map it to CPG Co.'s SKU nomenclature, scheme constructs and margin vocabulary. The base for this STT was a 0.6B-parameter Nvidia Parakeet model — our fine-tuned version delivered a lower word error rate (WER) than leading frontier speech models, at low latency.

Specialized SLMs deployed inside Palantir AIP

Transcripts then enter data pipelines built on Palantir AIP, where small language models — doing work that rules and regexes cannot — classify each exchange against a taxonomy of product perception, trade issues and competitive movement. They score retailer sentiment and extract typed entities (SKU, scheme, competitor, objection), which are written back into the ontology as linked objects. The same layer composes the rep's start-of-day briefing and end-of-day summary from that structured state, ensuring that the recommendations received by DSRs are always in sync with the live state of the data ontology.

The ontology is the substrate powering real-time recommendations

Hyde and Bain & Co. modelled the customer's operating world (DSRs, sales routes, outlets, SKUs, and the data from the pipeline described above) as a live ontology in AIP, where the entities are linked objects rather than rows in disconnected tables. This does two things. First, linked objects enable fast retrieval, allowing voice agents to work out the answer to unanticipated questions by walking the ontology, instead of serving pre-computed responses. Second, all processed transcripts are written back to the data ontology, creating a continuously improving recommendation loop: something learned in one visit powers recommendations for the next visit, even a fortnight later, and stays in the system when the rep moves on.

Small models enable economical voice AI at 50 million+ minutes a year

At CPG Co.'s volume, the running costs of this program using off-the-shelf models (per-minute audio pricing and per-token frontier context windows) are uneconomical. Our specialist models ensure that each conversation turn is a narrow query against a small context rather than a broad one against a large prompt. These smaller models cost a fraction of what frontier models cost to serve. Owning all layers of the AI stack — a custom STT, a custom LLM, and a custom TTS — enables us to keep the per-minute economics under control at scale.

Evals make recommendations reliable

Reliability at this scale needs to be measured rather than assumed. Every recommendation is scored against a golden dataset constructed from CPG Co.'s own data, with outliers flagged for review. The eval harness encodes what "good" means in CPG Co.'s commercial terms and is versioned with the system, which turns quality into a tracked regression surface rather than a launch-day claim.

What does this give CPG Co.?

The conversation that was previously discarded now resolves into a live, outlet-level map of the market, and it lands differently at each level of the organization. Reps get the pitch most likely to work in the store they are about to enter, ahead of the visit and in their own language. Territory sales executives see how their teams are actually selling, enabling them to provide evidence-based coaching to DSRs. Head office reads product perception, trade issues and competitive movement first-hand rather than renting them from a panel two months late. Because every layer queries the same ontology, everyone is reading the same live state rather than their own copy of it. The data corpus created is entirely owned by CPG Co., ensuring the resulting intelligence layer is a sovereign asset for the company.

What did the pilot prove?

Saathi was tested in selected cities against a control group, with store-level sales uplift as the key variable. Stores served by Saathi-equipped reps delivered a 1,200bps average revenue uplift versus control, comfortably ahead of the benchmark the teams had set. That was enough to turn a pilot into a national decision: rollout is planned across nearly 2 million stores, driving a $50M+ revenue opportunity above planned growth, with beat design, promotion management and structured upsell using the same pillars underpinning Saathi.

That is the work we care about at Hyde: taking frontier open-weight models and post-training them to solve a business objective and meeting a business where it operates. We’re proud to have deployed frontier intelligence to level up the frontline sales force.