Three months after the cleanup, the mess is back
In March you spent two weeks on the contact database. You merged 1,400 duplicates, filled in missing company names, standardized the state field so it stopped containing 'CA', 'Calif', 'California', and one brave 'Cali'. You exported the clean file, loaded it into your platform, and told everyone the list was finally trustworthy.
It's June. A rep just emailed a prospect named 'test test' at a company called 'asdf'. Marketing pulled a regional segment and got 340 records with no state at all. Someone on the sales team has been typing phone numbers as '(555) 123-4567' while the webform writes '5551234567', so your dialer integration is choking on half of them.
Nothing broke. People just kept working. Every form, import, trade show badge scan, and quick manual entry added records that nobody had agreed on the shape of. The cleanup fixed the past. It said nothing about the future.
That gap is what a data quality SLA closes. Not a data contract enforced by engineers at a schema level, which is what most writing on this topic assumes you have. A one-page agreement between the humans who create records, written in plain English, with names and dates on it.
Why decay is a process problem, not a people problem
Data degrades at a predictable rate because entry happens at a predictable rate. If your team adds 500 records a month and 8 percent of them are incomplete or malformed, you're generating 40 bad records a month without anyone being careless. Add a quarter of churn in job titles and email addresses and the decay compounds.
The reason it feels personal is that the symptoms show up as somebody's name. The rep who never fills in industry. The contractor who imported a list with the columns shifted one to the left. But those are predictable outcomes of a system where nobody ever wrote down what a complete record looks like.
So the enemy isn't the rep. It's the absence of a shared definition. Three teams touch the same contact object with three different ideas of what it's for. Sales wants a phone number. Marketing wants a consent flag and a source. Ops wants a billing country that matches a real country.
Our post on [building a data quality culture](/blog/building-a-data-quality-culture) covers the behavior side, and the [data governance for small teams](/blog/data-governance-for-small-teams) piece covers the decision rights. This article produces the artifact both of them describe: the actual page you print, sign, and pin above the quarterly review.
What a data quality SLA actually covers
Strip away the enterprise vocabulary and a data quality SLA answers four questions. What makes a record valid. What format each field takes. How recently it has to have been verified. Who fixes it when it fails, and by when.
Required fields come first because they're the easiest to argue about and the easiest to settle. Not every field is required. Resist the urge to mark twelve fields mandatory, because the result is reps typing 'n/a' into eight of them. Pick the three to five fields that actually block downstream work, and be specific about which object they apply to.
Formats are where most of the day-to-day friction lives. State codes, country codes, phone number shape, how you write company suffixes (Inc, Inc., Incorporated), date order, currency. Each of these is a one-line decision that saves an hour a month forever.
Freshness is the one teams skip. A contact record with a verified email from 2021 is not the same asset as one verified last quarter, even though both look identical in the interface. Set a staleness window per record type: open opportunities get checked more often than a dormant newsletter segment.
Ownership is the part that makes it an SLA instead of a style guide. Each rule needs a named role responsible for fixing failures and a response window. 'Sales ops, five business days' is an SLA. 'We should all keep the data clean' is a wish.
Writing the one-page agreement
Get sales, marketing, and ops in a room for 45 minutes. One person drives the doc on screen. The goal is not a perfect policy, it's a signed page by the end of the meeting.
Start by listing the five reports or workflows that break most often. Regional segmentation, lead routing, renewal reminders, the monthly revenue roll-up, the abandoned cart flow. Work backward from each one to the fields it depends on. That list is your required field set, and it's defensible, because every field on it traces to something a team actually uses.
Then go field by field and write the rule in one sentence a new hire could follow. 'State is the two-letter USPS code for US records and left blank for non-US.' 'Phone is digits only with a plus and country code.' 'Lead source must be one of the seven picklist values, and we do not add new ones without ops approval.' If a rule takes a paragraph, the rule is too complicated and people will route around it.
Assign an owner per field group rather than per field. Marketing owns source, consent, and campaign attribution. Sales owns title, phone, and account linkage. Ops owns billing address, currency, and anything with a tax implication. Three owners is enough. Six owners means nobody owns anything.
Keep it to one page. If it spills onto a second, you've written a reference manual, and reference manuals don't get read.
Setting thresholds people will actually hold
The fastest way to kill an SLA is to set the bar at 100 percent. The first week someone misses it, the number becomes a joke, and from then on everybody ignores the dashboard.
Measure where you are right now, then set the target one honest step above it. If 72 percent of new contacts this month arrived with a valid email and a source, make the target 85 percent for next quarter. You can tighten it in six months. A target you hit and then raise builds more momentum than a target you miss and then quietly delete.
Thresholds also need a scope. Apply them to records created in the current period, not to the entire historical database. Nobody can be accountable for a 2019 import. Holding the line on new records is the thing that stops decay, and it's the only measurement your team has real control over.
This is where a single composite number helps. CleanSmart's Clarity Score weights four dimensions, Completeness, Consistency, Uniqueness, and Validity, along with contact verification results, into one figure you can write into the agreement. It's recalculated after each cleaning run, which takes a few minutes, so the number in your SLA reflects your most recent run rather than a moving target you have to interpret live.
A reasonable first SLA has three numbers in it and no more: a Clarity Score floor for the database overall, a completeness target for newly created records, and a maximum duplicate rate. Three numbers fit in a head. Fifteen fit in a spreadsheet nobody opens.
Enforcement at intake versus detection after the fact
There are two places to stop bad data, and you need both. Intake controls prevent malformed records from being created. Detection catches what slips through anyway, including everything that arrives by import, integration, or a well-meaning colleague pasting 600 rows at once.
Intake controls are free and you already have them. They're unglamorous and they work:
- Turn free-text fields into picklists wherever the valid set is finite. Lead source, industry, region, lifecycle stage.
- Mark your three to five required fields as actually required in the form and the record layout, not just in the doc.
- Add format hints and input masks on webforms so phone and postal code arrive in one shape.
- Kill the fields nobody uses. Every unused field is a place for garbage to accumulate quietly.
- Give imports a checklist and one owner. Most catastrophic data events are one bad CSV, not a thousand bad keystrokes.
What detection looks like without a data team
Detection is the half people skip, because historically it meant someone writing queries. You don't need that. You need a scheduled pass that reads your records, compares them against the rules in your SLA, and tells you what failed.
In CleanSmart, DataBridge connectors read records from Mailchimp, Shopify, Klaviyo, HubSpot, or Salesforce so you're testing live data rather than a stale export someone saved to their desktop. SmartMatch finds the duplicates that intake controls can never prevent, because the same person signs up twice with two addresses and no form validation will ever catch that. AutoFormat standardizes the state codes and phone shapes back to the format your SLA names. SmartFill closes gaps where the missing value can be inferred from data you already hold. LogicGuard flags the anomalies: the order with a quantity of 4,000, the birth year of 1899, the billing country that doesn't match the currency.
The flow is import, clean, review, export. You review what the tools propose before anything is final, then take a CSV or JSON export plus a change log back into your platform yourself. Nothing changes in your systems without you putting it there, which matters a great deal when the whole point of the exercise is trust.
Practically: run the pass monthly, on the same day, before the review meeting. The output is your SLA scorecard. No queries, no engineer, no arguing from memory about whether things got better.
The monthly review ritual
An SLA without a standing meeting is a document. Thirty minutes a month, same slot, three people. Sales ops, marketing ops, whoever owns billing data.
Open with the numbers from the most recent cleaning run: Clarity Score, new-record completeness, duplicate rate. Compare to last month. Resist the urge to discuss anything else until those three are on the screen.
Then spend twenty minutes on the single largest failure category. Not all of them. One. If 180 records came in without a lead source, the question isn't who to blame, it's which form or import or integration produced them. Usually it's one source producing most of the mess, and fixing that source is a ten-minute change that holds for a year.
Close by updating the SLA itself. Rules that nobody followed get rewritten or deleted. Targets that have been hit three months running get raised by a notch. Note the date of the change at the bottom of the page. An SLA that never changes is an SLA nobody is reading.
Send the page, not the minutes. The one-page doc with the three numbers and the date is the entire record of the meeting.
A copyable SLA template
Copy this into a doc, fill in the brackets, delete what doesn't apply, and get three signatures. Keep it to one page.
Nothing here requires a tool. It requires that someone decides and that everyone can see what was decided.
- Title: [Company] Data Quality Agreement. Version [1.0]. Effective [date]. Next review [date].
- Scope: Applies to [contact and account records] in [platform]. Measured on records created or modified in the current month.
- Required fields (contact): email, first name, last name, lead source, [country]. A record missing any of these is incomplete.
- Required fields (account): legal name, billing country, currency, owner.
- Format rules: email lowercase, no leading or trailing spaces. US state as two-letter code. Country as two-letter ISO code. Phone as digits with plus and country code. Company name without suffix punctuation (Inc, not Inc.). Dates as YYYY-MM-DD.
- Picklists: lead source, industry, lifecycle stage, and region are picklists. New values require approval from [ops owner].
- Freshness: contacts on open opportunities verified within [90] days. Marketing-active contacts verified within [180] days. Dormant contacts flagged for review at [12] months.
- Uniqueness: one contact record per email address. One account record per legal entity. Duplicate rate target below [2] percent.
- Ownership: marketing owns source, consent, campaign fields. Sales owns title, phone, account linkage. Ops owns billing address, currency, tax fields.
- Response times: records flagged as invalid are corrected by the owning team within [5] business days. Anomalies affecting billing within [1] business day.
- Targets this quarter: Clarity Score at or above [78]. New-record completeness at or above [85] percent. Duplicate rate at or below [2] percent.
- Detection: a cleaning run is performed on the [first Monday] of each month. Results reviewed at the monthly data meeting.
- Signed: [sales ops], [marketing ops], [operations].
Make the enforcement half automatic
The writing part of a data quality SLA takes an afternoon. The hard part is the twelfth month, when the person who championed it has moved on and nobody is checking anything.
Automate the checking and the SLA outlives its author. Set a monthly cleaning run, let LogicGuard flag the records that break your rules, and let the scorecard arrive whether or not anyone remembers to ask for it. Your meeting then has an agenda it didn't have to produce. Your targets have evidence behind them. The three numbers on your one page stop being aspirational.
You cleaned the database once. The SLA is what keeps you from doing it again next March. Try a cleaning run on your own data in the [self-serve product demo](/demo) and see what your current records would score against the rules you're about to write.
Frequently Asked Questions
How is a data quality SLA different from a data contract?
A data contract is usually a technical specification enforced in code by engineers at the point where systems exchange data. A data quality SLA is the same idea written for people: plain-English rules about fields, formats, freshness, and owners, enforced through form settings, picklists, and a scheduled check. If you don't have a data team, the SLA is the version that actually gets used.
What thresholds should we start with?
Measure your current numbers first, then set targets one step above them rather than at perfection. A good starting set is three numbers only: a Clarity Score floor, a completeness target for newly created records, and a maximum duplicate rate. Raise them once you've hit them three months running.
Does CleanSmart apply the fixes in our CRM for us?
No. DataBridge connectors read records from Mailchimp, Shopify, Klaviyo, HubSpot, and Salesforce so you can clean and review them, then you get a CSV or JSON export plus a change log to bring back into your platform yourself. You stay in control of what changes in your systems.