B

Balsam Solutions

Bloom your business

Platforms

Pricing

Industries

Work

About

EmailClient login
07 · Build

AI that does a job in your business, not a demo.

We do not sell AI. We put it inside work you already do: the inbox that needs a first draft, the invoices that need their line items pulled out, the website visitor who needs a straight answer at 11pm, the call that needs a summary before the next one. Built on models you can swap and data that stays yours.

See pricing

Measured

every assistant ships with a before-and-after number

Swappable

Claude, OpenAI or a local model, changed in config, not a rewrite

$40–$200/mo

typical model cost, paid directly to the provider

What is in it

  • AI customer support

  • AI website chatbot

  • AI lead qualification

  • AI email drafting

  • AI document extraction

  • AI invoice processing

  • AI appointment assistant

  • AI knowledge base

  • AI-powered search

  • AI sales assistant

  • AI call summaries

  • AI-generated marketing content

  • AI internal assistant

Before and after

Four jobs, before and after an assistant is inside them.

The after column says what the assistant reads, what it is allowed to do on its own, and where a person steps in. How much it changed is measured in your business, not claimed on this page.

  1. A support question at 11pm

    A visitor on your website, after the last person has gone home.

    Today

    The chat widget asks for an email and promises a reply in the morning. The visitor wanted to know whether you cover their ZIP and whether Thursday works. By 8am they have booked with whoever answered.

    With the assistant

    The assistant answers from your service list, your areas, your prices and your policies, and nothing else. If the question is inside that, it answers and offers the next open slot from your booking calendar. If it is not, it says so, hands the visitor to the answering desk or leaves a text thread for the morning, and attaches the transcript. Every conversation is logged for the monthly review.

    What we measure

    After-hours conversations that end in a booking or a named next step.

  2. A vendor invoice

    A PDF, or a phone photo of a paper invoice, in the office inbox.

    Today

    Someone opens it, works out which property it belongs to, retypes the line items into QuickBooks and hopes the total matches. It waits in the inbox until someone has twenty minutes.

    With the assistant

    The invoice is read when it lands: line items, totals, dates, vendor. The assistant matches the vendor and the property against your records and drafts the bill in QuickBooks. Nothing posts. The draft waits in an approval queue with the original beside it, and a total that does not match its line items holds the draft with a note. A person approves, edits or rejects each one.

    What we measure

    Time from inbox to approved bill, and how many drafts a person had to correct.

  3. A sales call

    Forty minutes with a prospect, and the next call starts.

    Today

    What they asked for, what was promised and when to follow up live in the salesperson’s head, then in a CRM note written from memory on Friday, if it gets written at all.

    With the assistant

    The call is recorded with a notice at the start, transcribed, and summarised into a fixed shape: what they need, what was quoted, the objections, the agreed next step and its date. The summary and the follow-up task are written into the CRM record with the transcript linked, and the follow-up lands on the calendar. The salesperson reads it before the next call instead of writing it after this one.

    What we measure

    Calls with a next step and a date in the CRM within the hour.

  4. “Where is the SOP?”

    A new hire, in the group chat, for the fourth time this week.

    Today

    Three people answer, two of them differently. The document is in a Drive folder called FINAL v3, last touched by someone who left, and the copy on the shared drive is older than the one in the email thread.

    With the assistant

    The assistant has read the Drive, the SharePoint, the tickets and the policies, with each person’s permissions respected. It answers in plain English and cites the document and the section it came from, so the answer can be checked in one click. When it finds nothing, it says so instead of guessing, and the question is logged: the questions that come back empty are the list of SOPs that need writing.

    What we measure

    Questions answered with a citation, and the ones that came back empty.

Every assistant ships with a before-and-after number. The metric is agreed before we build, recorded before launch and every month after, and it lives in your account, not on our website. That is why this page does not quote one.

The only number we quote is the one your own dashboard shows.

Invoice processing for a property manager

Vendor invoices arrive as PDFs and phone photos. The assistant extracts the line items, matches the property, and drafts the QuickBooks bill for a human to approve.

  1. 1

    Invoice lands in a shared inbox

  2. 2

    Line items and totals extracted

  3. 3

    Property and vendor matched

  4. 4

    Bill drafted in QuickBooks, flagged for approval

Three ways to buy it

the flagship

AI Support, Chat & Lead Qualification

A website assistant that answers real questions, qualifies leads and books appointments at 11pm.

Measured

every assistant ships with a before-and-after number

From $3,500

AI Document & Invoice Processing

Invoices, receipts, forms and call recordings turned into structured data a human approves.

Human approves

the assistant drafts; nothing posts to the books without a person

From $4,500

AI Knowledge Base, Search & Internal Assistant

An assistant that has read your SOPs, contracts and history, and answers your team in plain English.

Cited

every answer links to the document it came from

From $6,000

How we keep it honest

Four rules under every assistant we build.

AI is easy to demo and hard to trust. These are the four things we do the same way on every job, whatever the task and whichever model is under it.

  1. Grounded in your data

    The assistant answers from your documents, your price list, your policies and your systems, and nothing else. Every answer carries the document it came from. When the answer is not in there, it says so, because a confident guess is worse than no answer.

    Sample exchange

    Visitor

    Do you cover 78745, and can someone come Thursday?

    Assistant

    Yes, 78745 is inside the south Austin service area. The first open Thursday slot is 8am to 10am. Want me to hold it and text you a confirmation?

    Sources:

    Service areas
    Booking calendar

  2. A person approves anything that writes

    Drafts, not actions. A bill, a CRM update, a reply to a customer: the assistant prepares it and a person approves, edits or rejects it from a queue. Nothing posts to the books without an approval row, and the row says who approved it and when.

    Sample queue

    • Bill draft · plumbing supply · Unit 4B

      Awaiting approval

    • Bill draft · electrical sub · Oak St job

      Held: total does not match line items

    • CRM note · Tuesday 2pm call

      Approved by the office manager

  3. The model is a setting, not the architecture

    Claude by default; OpenAI, or a model running on hardware you own, when the job or the budget says so. The model name is one line in a config file. Swapping it is a change to that line and a run of the eval set, not a rebuild, and you will not hear the word “migration”.

    assistant.config

    chat       = claude-sonnet-5extraction = claude-sonnet-5summaries  = claude-haiku-4-5-20251001# change the line, run the evals, done

  4. Nothing you give us trains anyone’s model

    We use the API tiers whose terms say your prompts and documents are not used for training. Your documents stay in storage in your name and the assistant reads them there. The provider account is yours too, so the usage bill and the data terms are between you and the provider, with us added as an admin.

    In your name

    • The provider account, and its bill

    • The storage the documents live in

    • The log of every conversation

    • The eval set and its results

If a job needs a judgement call on every case, an assistant does not remove the judgement; it queues it. We will say so on the first call, and sometimes the honest answer is a person and a checklist.

Pricing

From $3,500

Website chatbot or lead qualifier from $3,500. Document and invoice extraction from $4,500. Internal assistant over your own documents from $6,000.

Model usage billed to your own provider account at cost. Monitoring and prompt upkeep under a managed plan.

Add to a packageAll project rates

Questions

Whichever fits the job and the budget. We default to Claude for anything that reads documents or talks to customers, and we build so the model is a configuration value rather than an architecture decision.

Platforms we use here

HubSpot

Twilio

AWS

Google Workspace

QuickBooks

Microsoft 365

ServiceTitan

Postgres

Book the 30 minutes. Worst case, you leave with a plan you can hand to someone else.

No deck. No pitch. You talk about what your week actually looks like and we write down what we would fix first.

pst@balsamsolutions.com

The answering desk never closes. Nights, weekends, holidays, storm weeks.

Email